2 * This file is part of FFmpeg.
4 * FFmpeg is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU Lesser General Public
6 * License as published by the Free Software Foundation; either
7 * version 2.1 of the License, or (at your option) any later version.
9 * FFmpeg is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * Lesser General Public License for more details.
14 * You should have received a copy of the GNU Lesser General Public
15 * License along with FFmpeg; if not, write to the Free Software
16 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
18 * Copyright (C) 2000-2008, Intel Corporation, all rights reserved.
19 * Copyright (C) 2009, Willow Garage Inc., all rights reserved.
20 * Copyright (C) 2013, OpenCV Foundation, all rights reserved.
21 * Third party copyrights are property of their respective owners.
23 * Redistribution and use in source and binary forms, with or without modification,
24 * are permitted provided that the following conditions are met:
26 * * Redistribution's of source code must retain the above copyright notice,
27 * this list of conditions and the following disclaimer.
29 * * Redistribution's in binary form must reproduce the above copyright notice,
30 * this list of conditions and the following disclaimer in the documentation
31 * and/or other materials provided with the distribution.
33 * * The name of the copyright holders may not be used to endorse or promote products
34 * derived from this software without specific prior written permission.
36 * This software is provided by the copyright holders and contributors "as is" and
37 * any express or implied warranties, including, but not limited to, the implied
38 * warranties of merchantability and fitness for a particular purpose are disclaimed.
39 * In no event shall the Intel Corporation or contributors be liable for any direct,
40 * indirect, incidental, special, exemplary, or consequential damages
41 * (including, but not limited to, procurement of substitute goods or services;
42 * loss of use, data, or profits; or business interruption) however caused
43 * and on any theory of liability, whether in contract, strict liability,
44 * or tort (including negligence or otherwise) arising in any way out of
45 * the use of this software, even if advised of the possibility of such damage.
50 #include <libavutil/lfg.h>
51 #include "libavutil/opt.h"
52 #include "libavutil/imgutils.h"
53 #include "libavutil/mem.h"
54 #include "libavutil/fifo.h"
55 #include "libavutil/common.h"
56 #include "libavutil/avassert.h"
57 #include "libavutil/pixfmt.h"
59 #include "framequeue.h"
61 #include "transform.h"
65 #include "opencl_source.h"
69 This filter matches feature points between frames (dealing with outliers) and then
70 uses the matches to estimate an affine transform between frames. This transform is
71 decomposed into various values (translation, scale, rotation) and the values are
72 summed relative to the start of the video to obtain on absolute camera position
73 for each frame. This "camera path" is then smoothed via a gaussian filter, resulting
74 in a new path that is turned back into an affine transform and applied to each
79 All of the work to extract motion data from frames occurs in queue_frame. Motion data
80 is buffered in a smoothing window, so queue_frame simply computes the absolute camera
81 positions and places them in ringbuffers.
83 filter_frame is responsible for looking at the absolute camera positions currently
84 in the ringbuffers, applying the gaussian filter, and then transforming the frames.
87 // Number of bits for BRIEF descriptors
89 // Size of the patch from which a BRIEF descriptor is extracted
90 // This is the size used in OpenCV
91 #define BRIEF_PATCH_SIZE 31
92 #define BRIEF_PATCH_SIZE_HALF (BRIEF_PATCH_SIZE / 2)
94 #define MATCHES_CONTIG_SIZE 2000
96 #define ROUNDED_UP_DIV(a, b) ((a + (b - 1)) / b)
98 typedef struct PointPair {
105 typedef struct MotionVector {
107 // Used to mark vectors as potential outliers
108 cl_int should_consider;
111 // Denotes the indices for the different types of motion in the ringbuffers array
112 enum RingbufferIndices {
119 // Should always be last
123 // Struct that holds data for drawing point match debug data
124 typedef struct DebugMatches {
125 MotionVector *matches;
126 // The points used to calculate the affine transform for a frame
127 MotionVector model_matches[3];
130 // For cases where we couldn't calculate a model
131 int num_model_matches;
134 // Groups together the ringbuffers that store absolute distortion / position values
136 typedef struct AbsoluteFrameMotion {
137 // Array with the various ringbuffers, indexed via the RingbufferIndices enum
138 AVFifoBuffer *ringbuffers[RingbufCount];
140 // Offset to get to the current frame being processed
142 int curr_frame_offset;
143 // Keeps track of where the start and end of contiguous motion data is (to
144 // deal with cases where no motion data is found between two frames)
145 int data_start_offset;
148 AVFifoBuffer *debug_matches;
149 } AbsoluteFrameMotion;
151 // Takes care of freeing the arrays within the DebugMatches inside of the
152 // debug_matches ringbuffer and then freeing the buffer itself.
153 static void free_debug_matches(AbsoluteFrameMotion *afm) {
156 if (!afm->debug_matches) {
160 while (av_fifo_size(afm->debug_matches) > 0) {
161 av_fifo_generic_read(
164 sizeof(DebugMatches),
168 av_freep(&dm.matches);
171 av_fifo_freep(&afm->debug_matches);
174 // Stores the translation, scale, rotation, and skew deltas between two frames
175 typedef struct FrameDelta {
176 cl_float2 translation;
182 typedef struct SimilarityMatrix {
183 // The 2x3 similarity matrix
187 typedef struct CropInfo {
188 // The top left corner of the bounding box for the crop
190 // The bottom right corner of the bounding box for the crop
191 cl_float2 bottom_right;
194 // Returned from function that determines start and end values for iteration
195 // around the current frame in a ringbuffer
196 typedef struct IterIndices {
201 typedef struct DeshakeOpenCLContext {
202 OpenCLFilterContext ocf;
203 // Whether or not the above `OpenCLFilterContext` has been initialized
206 // These variables are used in the activate callback
210 // State for random number generation
213 // FIFO frame queue used to buffer future frames for processing
215 // Ringbuffers for frame positions
216 AbsoluteFrameMotion abs_motion;
218 // The number of frames' motion to consider before and after the frame we are
221 // The number of the frame we are currently processing
224 // Stores a 1d array of normalised gaussian kernel values for convolution
227 // Buffer for error values used in RANSAC code
230 // Information regarding how to crop the smoothed luminance (or RGB) planes
232 // Information regarding how to crop the smoothed chroma planes
235 // Whether or not we are processing YUV input (as oppposed to RGB)
237 // The underlying format of the hardware surfaces
240 // Buffer to copy `matches` into for the CPU to work with
241 MotionVector *matches_host;
242 MotionVector *matches_contig_host;
244 MotionVector *inliers;
246 cl_command_queue command_queue;
247 cl_kernel kernel_grayscale;
248 cl_kernel kernel_harris_response;
249 cl_kernel kernel_refine_features;
250 cl_kernel kernel_brief_descriptors;
251 cl_kernel kernel_match_descriptors;
252 cl_kernel kernel_transform;
253 cl_kernel kernel_crop_upscale;
255 // Stores a frame converted to grayscale
257 // Stores the harris response for a frame (measure of "cornerness" for each pixel)
260 // Detected features after non-maximum suppression and sub-pixel refinement
261 cl_mem refined_features;
262 // Saved from the previous frame
263 cl_mem prev_refined_features;
265 // BRIEF sampling pattern that is randomly initialized
266 cl_mem brief_pattern;
267 // Feature point descriptors for the current frame
269 // Feature point descriptors for the previous frame
270 cl_mem prev_descriptors;
271 // Vectors between points in current and previous frame
273 cl_mem matches_contig;
274 // Holds the matrix to transform luminance (or RGB) with
276 // Holds the matrix to transform chroma with
279 // Configurable options
285 // Whether or not feature points should be refined at a sub-pixel level
286 cl_int refine_features;
287 // If the user sets a value other than the default, 0, this percentage is
288 // translated into a sigma value ranging from 0.5 to 40.0
289 float smooth_percent;
290 // This number is multiplied by the video frame rate to determine the size
291 // of the smooth window
292 float smooth_window_multiplier;
296 cl_kernel kernel_draw_debug_info;
297 cl_mem debug_matches;
298 cl_mem debug_model_matches;
300 // These store the total time spent executing the different kernels in nanoseconds
301 unsigned long long grayscale_time;
302 unsigned long long harris_response_time;
303 unsigned long long refine_features_time;
304 unsigned long long brief_descriptors_time;
305 unsigned long long match_descriptors_time;
306 unsigned long long transform_time;
307 unsigned long long crop_upscale_time;
309 // Time spent copying matched features from the device to the host
310 unsigned long long read_buf_time;
311 } DeshakeOpenCLContext;
313 // Returns a random uniformly-distributed number in [low, high]
314 static int rand_in(int low, int high, AVLFG *alfg) {
315 return (av_lfg_get(alfg) % (high - low)) + low;
318 // Returns the average execution time for an event given the total time and the
319 // number of frames processed.
320 static double averaged_event_time_ms(unsigned long long total_time, int num_frames) {
321 return (double)total_time / (double)num_frames / 1000000.0;
324 // The following code is loosely ported from OpenCV
326 // Estimates affine transform from 3 point pairs
327 // model is a 2x3 matrix:
330 static void run_estimate_kernel(const MotionVector *point_pairs, double *model)
333 double x1 = point_pairs[0].p.p1.s[0];
334 double y1 = point_pairs[0].p.p1.s[1];
335 double x2 = point_pairs[1].p.p1.s[0];
336 double y2 = point_pairs[1].p.p1.s[1];
337 double x3 = point_pairs[2].p.p1.s[0];
338 double y3 = point_pairs[2].p.p1.s[1];
341 double X1 = point_pairs[0].p.p2.s[0];
342 double Y1 = point_pairs[0].p.p2.s[1];
343 double X2 = point_pairs[1].p.p2.s[0];
344 double Y2 = point_pairs[1].p.p2.s[1];
345 double X3 = point_pairs[2].p.p2.s[0];
346 double Y3 = point_pairs[2].p.p2.s[1];
348 double d = 1.0 / ( x1*(y2-y3) + x2*(y3-y1) + x3*(y1-y2) );
350 model[0] = d * ( X1*(y2-y3) + X2*(y3-y1) + X3*(y1-y2) );
351 model[1] = d * ( X1*(x3-x2) + X2*(x1-x3) + X3*(x2-x1) );
352 model[2] = d * ( X1*(x2*y3 - x3*y2) + X2*(x3*y1 - x1*y3) + X3*(x1*y2 - x2*y1) );
354 model[3] = d * ( Y1*(y2-y3) + Y2*(y3-y1) + Y3*(y1-y2) );
355 model[4] = d * ( Y1*(x3-x2) + Y2*(x1-x3) + Y3*(x2-x1) );
356 model[5] = d * ( Y1*(x2*y3 - x3*y2) + Y2*(x3*y1 - x1*y3) + Y3*(x1*y2 - x2*y1) );
359 // Checks that the 3 points in the given array are not collinear
360 static bool points_not_collinear(const cl_float2 **points)
364 for (j = 0; j < i; j++) {
365 double dx1 = points[j]->s[0] - points[i]->s[0];
366 double dy1 = points[j]->s[1] - points[i]->s[1];
368 for (k = 0; k < j; k++) {
369 double dx2 = points[k]->s[0] - points[i]->s[0];
370 double dy2 = points[k]->s[1] - points[i]->s[1];
372 // Assuming a 3840 x 2160 video with a point at (0, 0) and one at
373 // (3839, 2159), this prevents a third point from being within roughly
374 // 0.5 of a pixel of the line connecting the two on both axes
375 if (fabs(dx2*dy1 - dy2*dx1) <= 1.0) {
384 // Checks a subset of 3 point pairs to make sure that the points are not collinear
385 // and not too close to each other
386 static bool check_subset(const MotionVector *pairs_subset)
388 const cl_float2 *prev_points[] = {
389 &pairs_subset[0].p.p1,
390 &pairs_subset[1].p.p1,
391 &pairs_subset[2].p.p1
394 const cl_float2 *curr_points[] = {
395 &pairs_subset[0].p.p2,
396 &pairs_subset[1].p.p2,
397 &pairs_subset[2].p.p2
400 return points_not_collinear(prev_points) && points_not_collinear(curr_points);
403 // Selects a random subset of 3 points from point_pairs and places them in pairs_subset
404 static bool get_subset(
406 const MotionVector *point_pairs,
407 const int num_point_pairs,
408 MotionVector *pairs_subset,
412 int i = 0, j, iters = 0;
414 for (; iters < max_attempts; iters++) {
415 for (i = 0; i < 3 && iters < max_attempts;) {
419 idx_i = idx[i] = rand_in(0, num_point_pairs, alfg);
421 for (j = 0; j < i; j++) {
422 if (idx_i == idx[j]) {
432 pairs_subset[i] = point_pairs[idx[i]];
436 if (i == 3 && !check_subset(pairs_subset)) {
442 return i == 3 && iters < max_attempts;
445 // Computes the error for each of the given points based on the given model.
446 static void compute_error(
447 const MotionVector *point_pairs,
448 const int num_point_pairs,
452 double F0 = model[0], F1 = model[1], F2 = model[2];
453 double F3 = model[3], F4 = model[4], F5 = model[5];
455 for (int i = 0; i < num_point_pairs; i++) {
456 const cl_float2 *f = &point_pairs[i].p.p1;
457 const cl_float2 *t = &point_pairs[i].p.p2;
459 double a = F0*f->s[0] + F1*f->s[1] + F2 - t->s[0];
460 double b = F3*f->s[0] + F4*f->s[1] + F5 - t->s[1];
466 // Determines which of the given point matches are inliers for the given model
467 // based on the specified threshold.
469 // err must be an array of num_point_pairs length
470 static int find_inliers(
471 MotionVector *point_pairs,
472 const int num_point_pairs,
477 float t = (float)(thresh * thresh);
478 int i, n = num_point_pairs, num_inliers = 0;
480 compute_error(point_pairs, num_point_pairs, model, err);
482 for (i = 0; i < n; i++) {
485 point_pairs[i].should_consider = true;
488 point_pairs[i].should_consider = false;
495 // Determines the number of iterations required to achieve the desired confidence level.
497 // The equation used to determine the number of iterations to do is:
498 // 1 - confidence = (1 - inlier_probability^num_points)^num_iters
500 // Solving for num_iters:
502 // num_iters = log(1 - confidence) / log(1 - inlier_probability^num_points)
504 // A more in-depth explanation can be found at https://en.wikipedia.org/wiki/Random_sample_consensus
505 // under the 'Parameters' heading
506 static int ransac_update_num_iters(double confidence, double num_outliers, int max_iters)
510 confidence = av_clipd(confidence, 0.0, 1.0);
511 num_outliers = av_clipd(num_outliers, 0.0, 1.0);
513 // avoid inf's & nan's
514 num = FFMAX(1.0 - confidence, DBL_MIN);
515 denom = 1.0 - pow(1.0 - num_outliers, 3);
516 if (denom < DBL_MIN) {
523 return denom >= 0 || -num >= max_iters * (-denom) ? max_iters : (int)round(num / denom);
526 // Estimates an affine transform between the given pairs of points using RANdom
528 static bool estimate_affine_2d(
529 DeshakeOpenCLContext *deshake_ctx,
530 MotionVector *point_pairs,
531 DebugMatches *debug_matches,
532 const int num_point_pairs,
534 const double threshold,
536 const double confidence
539 double best_model[6], model[6];
540 MotionVector pairs_subset[3], best_pairs[3];
542 int iter, niters = FFMAX(max_iters, 1);
543 int good_count, max_good_count = 0;
545 // We need at least 3 points to build a model from
546 if (num_point_pairs < 3) {
548 } else if (num_point_pairs == 3) {
549 // There are only 3 points, so RANSAC doesn't apply here
550 run_estimate_kernel(point_pairs, model_out);
552 for (int i = 0; i < 3; ++i) {
553 point_pairs[i].should_consider = true;
559 for (iter = 0; iter < niters; ++iter) {
560 bool found = get_subset(&deshake_ctx->alfg, point_pairs, num_point_pairs, pairs_subset, 10000);
570 run_estimate_kernel(pairs_subset, model);
571 good_count = find_inliers(point_pairs, num_point_pairs, model, deshake_ctx->ransac_err, threshold);
573 if (good_count > FFMAX(max_good_count, 2)) {
574 for (int mi = 0; mi < 6; ++mi) {
575 best_model[mi] = model[mi];
578 for (int pi = 0; pi < 3; pi++) {
579 best_pairs[pi] = pairs_subset[pi];
582 max_good_count = good_count;
583 niters = ransac_update_num_iters(
585 (double)(num_point_pairs - good_count) / num_point_pairs,
591 if (max_good_count > 0) {
592 for (int mi = 0; mi < 6; ++mi) {
593 model_out[mi] = best_model[mi];
596 for (int pi = 0; pi < 3; ++pi) {
597 debug_matches->model_matches[pi] = best_pairs[pi];
599 debug_matches->num_model_matches = 3;
601 // Find the inliers again for the best model for debugging
602 find_inliers(point_pairs, num_point_pairs, best_model, deshake_ctx->ransac_err, threshold);
609 // "Wiggles" the first point in best_pairs around a tiny bit in order to decrease the
611 static void optimize_model(
612 DeshakeOpenCLContext *deshake_ctx,
613 MotionVector *best_pairs,
614 MotionVector *inliers,
615 const int num_inliers,
619 float move_x_val = 0.01;
620 float move_y_val = 0.01;
622 float old_move_x_val = 0;
624 int last_changed = 0;
626 for (int iters = 0; iters < 200; iters++) {
630 best_pairs[0].p.p2.s[0] += move_x_val;
632 best_pairs[0].p.p2.s[0] += move_y_val;
635 run_estimate_kernel(best_pairs, model);
636 compute_error(inliers, num_inliers, model, deshake_ctx->ransac_err);
638 for (int j = 0; j < num_inliers; j++) {
639 total_err += deshake_ctx->ransac_err[j];
642 if (total_err < best_err) {
643 for (int mi = 0; mi < 6; ++mi) {
644 model_out[mi] = model[mi];
647 best_err = total_err;
648 last_changed = iters;
652 best_pairs[0].p.p2.s[0] -= move_x_val;
654 best_pairs[0].p.p2.s[0] -= move_y_val;
657 if (iters - last_changed > 4) {
658 // We've already improved the model as much as we can
662 old_move_x_val = move_x_val;
670 if (old_move_x_val < 0) {
679 // Uses a process similar to that of RANSAC to find a transform that minimizes
680 // the total error for a set of point matches determined to be inliers
682 // (Pick random subsets, compute model, find total error, iterate until error
684 static bool minimize_error(
685 DeshakeOpenCLContext *deshake_ctx,
686 MotionVector *inliers,
687 DebugMatches *debug_matches,
688 const int num_inliers,
693 float best_err = FLT_MAX;
694 double best_model[6], model[6];
695 MotionVector pairs_subset[3], best_pairs[3];
697 for (int i = 0; i < max_iters; i++) {
699 bool found = get_subset(&deshake_ctx->alfg, inliers, num_inliers, pairs_subset, 10000);
709 run_estimate_kernel(pairs_subset, model);
710 compute_error(inliers, num_inliers, model, deshake_ctx->ransac_err);
712 for (int j = 0; j < num_inliers; j++) {
713 total_err += deshake_ctx->ransac_err[j];
716 if (total_err < best_err) {
717 for (int mi = 0; mi < 6; ++mi) {
718 best_model[mi] = model[mi];
721 for (int pi = 0; pi < 3; pi++) {
722 best_pairs[pi] = pairs_subset[pi];
725 best_err = total_err;
729 for (int mi = 0; mi < 6; ++mi) {
730 model_out[mi] = best_model[mi];
733 for (int pi = 0; pi < 3; ++pi) {
734 debug_matches->model_matches[pi] = best_pairs[pi];
736 debug_matches->num_model_matches = 3;
739 optimize_model(deshake_ctx, best_pairs, inliers, num_inliers, best_err, model_out);
743 // End code from OpenCV
745 // Decomposes a similarity matrix into translation, rotation, scale, and skew
747 // See http://frederic-wang.fr/decomposition-of-2d-transform-matrices.html
748 static FrameDelta decompose_transform(double *model)
758 double delta = a * d - b * c;
760 ret.translation.s[0] = e;
761 ret.translation.s[1] = f;
763 // This is the QR method
764 if (a != 0 || b != 0) {
765 double r = hypot(a, b);
767 ret.rotation = FFSIGN(b) * acos(a / r);
769 ret.scale.s[1] = delta / r;
770 ret.skew.s[0] = atan((a * c + b * d) / (r * r));
772 } else if (c != 0 || d != 0) {
773 double s = sqrt(c * c + d * d);
775 ret.rotation = M_PI / 2 - FFSIGN(d) * acos(-c / s);
776 ret.scale.s[0] = delta / s;
779 ret.skew.s[1] = atan((a * c + b * d) / (s * s));
780 } // otherwise there is only translation
785 // Move valid vectors from the 2d buffer into a 1d buffer where they are contiguous
786 static int make_vectors_contig(
787 DeshakeOpenCLContext *deshake_ctx,
793 for (int i = 0; i < size_y; ++i) {
794 for (int j = 0; j < size_x; ++j) {
795 MotionVector v = deshake_ctx->matches_host[j + i * size_x];
797 if (v.should_consider) {
798 deshake_ctx->matches_contig_host[num_vectors] = v;
802 // Make sure we do not exceed the amount of space we allocated for these vectors
803 if (num_vectors == MATCHES_CONTIG_SIZE - 1) {
811 // Returns the gaussian kernel value for the given x coordinate and sigma value
812 static float gaussian_for(int x, float sigma) {
813 return 1.0f / expf(((float)x * (float)x) / (2.0f * sigma * sigma));
816 // Makes a normalized gaussian kernel of the given length for the given sigma
817 // and places it in gauss_kernel
818 static void make_gauss_kernel(float *gauss_kernel, float length, float sigma)
821 int window_half = length / 2;
823 for (int i = 0; i < length; ++i) {
824 float val = gaussian_for(i - window_half, sigma);
827 gauss_kernel[i] = val;
830 // Normalize the gaussian values
831 for (int i = 0; i < length; ++i) {
832 gauss_kernel[i] /= gauss_sum;
836 // Returns indices to start and end iteration at in order to iterate over a window
837 // of length size centered at the current frame in a ringbuffer
839 // Always returns numbers that result in a window of length size, even if that
840 // means specifying negative indices or indices past the end of the values in the
841 // ringbuffers. Make sure you clip indices appropriately within your loop.
842 static IterIndices start_end_for(DeshakeOpenCLContext *deshake_ctx, int length) {
845 indices.start = deshake_ctx->abs_motion.curr_frame_offset - (length / 2);
846 indices.end = deshake_ctx->abs_motion.curr_frame_offset + (length / 2) + (length % 2);
851 // Sets val to the value in the given ringbuffer at the given offset, taking care of
852 // clipping the offset into the appropriate range
853 static void ringbuf_float_at(
854 DeshakeOpenCLContext *deshake_ctx,
855 AVFifoBuffer *values,
859 int clip_start, clip_end, offset_clipped;
860 if (deshake_ctx->abs_motion.data_end_offset != -1) {
861 clip_end = deshake_ctx->abs_motion.data_end_offset;
863 // This expression represents the last valid index in the buffer,
864 // which we use repeatedly at the end of the video.
865 clip_end = deshake_ctx->smooth_window - (av_fifo_space(values) / sizeof(float)) - 1;
868 if (deshake_ctx->abs_motion.data_start_offset != -1) {
869 clip_start = deshake_ctx->abs_motion.data_start_offset;
871 // Negative indices will occur at the start of the video, and we want
872 // them to be clipped to 0 in order to repeatedly use the position of
877 offset_clipped = av_clip(
883 av_fifo_generic_peek_at(
886 offset_clipped * sizeof(float),
892 // Returns smoothed current frame value of the given buffer of floats based on the
893 // given Gaussian kernel and its length (also the window length, centered around the
894 // current frame) and the "maximum value" of the motion.
896 // This "maximum value" should be the width / height of the image in the case of
897 // translation and an empirically chosen constant for rotation / scale.
899 // The sigma chosen to generate the final gaussian kernel with used to smooth the
900 // camera path is either hardcoded (set by user, deshake_ctx->smooth_percent) or
901 // adaptively chosen.
903 DeshakeOpenCLContext *deshake_ctx,
909 float new_large_s = 0, new_small_s = 0, new_best = 0, old, diff_between,
910 percent_of_max, inverted_percent;
911 IterIndices indices = start_end_for(deshake_ctx, length);
912 float large_sigma = 40.0f;
913 float small_sigma = 2.0f;
916 if (deshake_ctx->smooth_percent) {
917 best_sigma = (large_sigma - 0.5f) * deshake_ctx->smooth_percent + 0.5f;
919 // Strategy to adaptively smooth trajectory:
921 // 1. Smooth path with large and small sigma values
922 // 2. Take the absolute value of the difference between them
923 // 3. Get a percentage by putting the difference over the "max value"
924 // 4, Invert the percentage
925 // 5. Calculate a new sigma value weighted towards the larger sigma value
926 // 6. Determine final smoothed trajectory value using that sigma
928 make_gauss_kernel(gauss_kernel, length, large_sigma);
929 for (int i = indices.start, j = 0; i < indices.end; ++i, ++j) {
930 ringbuf_float_at(deshake_ctx, values, &old, i);
931 new_large_s += old * gauss_kernel[j];
934 make_gauss_kernel(gauss_kernel, length, small_sigma);
935 for (int i = indices.start, j = 0; i < indices.end; ++i, ++j) {
936 ringbuf_float_at(deshake_ctx, values, &old, i);
937 new_small_s += old * gauss_kernel[j];
940 diff_between = fabsf(new_large_s - new_small_s);
941 percent_of_max = diff_between / max_val;
942 inverted_percent = 1 - percent_of_max;
943 best_sigma = large_sigma * powf(inverted_percent, 40);
946 make_gauss_kernel(gauss_kernel, length, best_sigma);
947 for (int i = indices.start, j = 0; i < indices.end; ++i, ++j) {
948 ringbuf_float_at(deshake_ctx, values, &old, i);
949 new_best += old * gauss_kernel[j];
955 // Returns the position of the given point after the transform is applied
956 static cl_float2 transformed_point(float x, float y, float *transform) {
959 ret.s[0] = x * transform[0] + y * transform[1] + transform[2];
960 ret.s[1] = x * transform[3] + y * transform[4] + transform[5];
965 // Creates an affine transform that scales from the center of a frame
966 static void transform_center_scale(
977 float center_s_w, center_s_h;
988 center_s = transformed_point(center_w, center_h, matrix);
989 center_s_w = center_w - center_s.s[0];
990 center_s_h = center_h - center_s.s[1];
993 x_shift + center_s_w,
994 y_shift + center_s_h,
1002 // Determines the crop necessary to eliminate black borders from a smoothed frame
1003 // and updates target crop accordingly
1004 static void update_needed_crop(
1010 float new_width, new_height, adjusted_width, adjusted_height, adjusted_x, adjusted_y;
1012 cl_float2 top_left = transformed_point(0, 0, transform);
1013 cl_float2 top_right = transformed_point(frame_width, 0, transform);
1014 cl_float2 bottom_left = transformed_point(0, frame_height, transform);
1015 cl_float2 bottom_right = transformed_point(frame_width, frame_height, transform);
1016 float ar_h = frame_height / frame_width;
1017 float ar_w = frame_width / frame_height;
1019 if (crop->bottom_right.s[0] == 0) {
1020 // The crop hasn't been set to the original size of the plane
1021 crop->bottom_right.s[0] = frame_width;
1022 crop->bottom_right.s[1] = frame_height;
1025 crop->top_left.s[0] = FFMAX3(
1026 crop->top_left.s[0],
1031 crop->top_left.s[1] = FFMAX3(
1032 crop->top_left.s[1],
1037 crop->bottom_right.s[0] = FFMIN3(
1038 crop->bottom_right.s[0],
1043 crop->bottom_right.s[1] = FFMIN3(
1044 crop->bottom_right.s[1],
1049 // Make sure our potentially new bounding box has the same aspect ratio
1050 new_height = crop->bottom_right.s[1] - crop->top_left.s[1];
1051 new_width = crop->bottom_right.s[0] - crop->top_left.s[0];
1053 adjusted_width = new_height * ar_w;
1054 adjusted_x = crop->bottom_right.s[0] - adjusted_width;
1056 if (adjusted_x >= crop->top_left.s[0]) {
1057 crop->top_left.s[0] = adjusted_x;
1059 adjusted_height = new_width * ar_h;
1060 adjusted_y = crop->bottom_right.s[1] - adjusted_height;
1061 crop->top_left.s[1] = adjusted_y;
1065 static av_cold void deshake_opencl_uninit(AVFilterContext *avctx)
1067 DeshakeOpenCLContext *ctx = avctx->priv;
1070 for (int i = 0; i < RingbufCount; i++)
1071 av_fifo_freep(&ctx->abs_motion.ringbuffers[i]);
1074 free_debug_matches(&ctx->abs_motion);
1076 if (ctx->gauss_kernel)
1077 av_freep(&ctx->gauss_kernel);
1079 if (ctx->ransac_err)
1080 av_freep(&ctx->ransac_err);
1082 if (ctx->matches_host)
1083 av_freep(&ctx->matches_host);
1085 if (ctx->matches_contig_host)
1086 av_freep(&ctx->matches_contig_host);
1089 av_freep(&ctx->inliers);
1091 ff_framequeue_free(&ctx->fq);
1093 CL_RELEASE_KERNEL(ctx->kernel_grayscale);
1094 CL_RELEASE_KERNEL(ctx->kernel_harris_response);
1095 CL_RELEASE_KERNEL(ctx->kernel_refine_features);
1096 CL_RELEASE_KERNEL(ctx->kernel_brief_descriptors);
1097 CL_RELEASE_KERNEL(ctx->kernel_match_descriptors);
1098 CL_RELEASE_KERNEL(ctx->kernel_crop_upscale);
1100 CL_RELEASE_KERNEL(ctx->kernel_draw_debug_info);
1102 CL_RELEASE_QUEUE(ctx->command_queue);
1105 CL_RELEASE_MEMORY(ctx->grayscale);
1106 CL_RELEASE_MEMORY(ctx->harris_buf);
1107 CL_RELEASE_MEMORY(ctx->refined_features);
1108 CL_RELEASE_MEMORY(ctx->prev_refined_features);
1109 CL_RELEASE_MEMORY(ctx->brief_pattern);
1110 CL_RELEASE_MEMORY(ctx->descriptors);
1111 CL_RELEASE_MEMORY(ctx->prev_descriptors);
1112 CL_RELEASE_MEMORY(ctx->matches);
1113 CL_RELEASE_MEMORY(ctx->matches_contig);
1114 CL_RELEASE_MEMORY(ctx->transform_y);
1115 CL_RELEASE_MEMORY(ctx->transform_uv);
1116 if (ctx->debug_on) {
1117 CL_RELEASE_MEMORY(ctx->debug_matches);
1118 CL_RELEASE_MEMORY(ctx->debug_model_matches);
1121 ff_opencl_filter_uninit(avctx);
1124 static int deshake_opencl_init(AVFilterContext *avctx)
1126 DeshakeOpenCLContext *ctx = avctx->priv;
1127 AVFilterLink *outlink = avctx->outputs[0];
1128 AVFilterLink *inlink = avctx->inputs[0];
1129 // Pointer to the host-side pattern buffer to be initialized and then copied
1131 PointPair *pattern_host;
1134 cl_ulong8 zeroed_ulong8;
1135 FFFrameQueueGlobal fqg;
1136 cl_image_format grayscale_format;
1137 cl_image_desc grayscale_desc;
1138 cl_command_queue_properties queue_props;
1140 const enum AVPixelFormat disallowed_formats[14] = {
1144 AV_PIX_FMT_GBRP10BE,
1145 AV_PIX_FMT_GBRP10LE,
1146 AV_PIX_FMT_GBRP16BE,
1147 AV_PIX_FMT_GBRP16LE,
1149 AV_PIX_FMT_GBRAP16BE,
1150 AV_PIX_FMT_GBRAP16LE,
1151 AV_PIX_FMT_GBRAP12BE,
1152 AV_PIX_FMT_GBRAP12LE,
1153 AV_PIX_FMT_GBRAP10BE,
1154 AV_PIX_FMT_GBRAP10LE
1157 // Number of elements for an array
1158 const int image_grid_32 = ROUNDED_UP_DIV(outlink->h, 32) * ROUNDED_UP_DIV(outlink->w, 32);
1160 const int descriptor_buf_size = image_grid_32 * (BREIFN / 8);
1161 const int features_buf_size = image_grid_32 * sizeof(cl_float2);
1163 const AVHWFramesContext *hw_frames_ctx = (AVHWFramesContext*)inlink->hw_frames_ctx->data;
1164 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(hw_frames_ctx->sw_format);
1166 av_assert0(hw_frames_ctx);
1169 ff_framequeue_global_init(&fqg);
1170 ff_framequeue_init(&ctx->fq, &fqg);
1172 ctx->smooth_window = (int)(av_q2d(avctx->inputs[0]->frame_rate) * ctx->smooth_window_multiplier);
1173 ctx->curr_frame = 0;
1175 memset(&zeroed_ulong8, 0, sizeof(cl_ulong8));
1177 ctx->gauss_kernel = av_malloc_array(ctx->smooth_window, sizeof(float));
1178 if (!ctx->gauss_kernel) {
1179 err = AVERROR(ENOMEM);
1183 ctx->ransac_err = av_malloc_array(MATCHES_CONTIG_SIZE, sizeof(float));
1184 if (!ctx->ransac_err) {
1185 err = AVERROR(ENOMEM);
1189 for (int i = 0; i < RingbufCount; i++) {
1190 ctx->abs_motion.ringbuffers[i] = av_fifo_alloc_array(
1195 if (!ctx->abs_motion.ringbuffers[i]) {
1196 err = AVERROR(ENOMEM);
1201 if (ctx->debug_on) {
1202 ctx->abs_motion.debug_matches = av_fifo_alloc_array(
1203 ctx->smooth_window / 2,
1204 sizeof(DebugMatches)
1207 if (!ctx->abs_motion.debug_matches) {
1208 err = AVERROR(ENOMEM);
1213 ctx->abs_motion.curr_frame_offset = 0;
1214 ctx->abs_motion.data_start_offset = -1;
1215 ctx->abs_motion.data_end_offset = -1;
1217 pattern_host = av_malloc_array(BREIFN, sizeof(PointPair));
1218 if (!pattern_host) {
1219 err = AVERROR(ENOMEM);
1223 ctx->matches_host = av_malloc_array(image_grid_32, sizeof(MotionVector));
1224 if (!ctx->matches_host) {
1225 err = AVERROR(ENOMEM);
1229 ctx->matches_contig_host = av_malloc_array(MATCHES_CONTIG_SIZE, sizeof(MotionVector));
1230 if (!ctx->matches_contig_host) {
1231 err = AVERROR(ENOMEM);
1235 ctx->inliers = av_malloc_array(MATCHES_CONTIG_SIZE, sizeof(MotionVector));
1236 if (!ctx->inliers) {
1237 err = AVERROR(ENOMEM);
1241 // Initializing the patch pattern for building BREIF descriptors with
1242 av_lfg_init(&ctx->alfg, 234342424);
1243 for (int i = 0; i < BREIFN; ++i) {
1246 for (int j = 0; j < 2; ++j) {
1247 pair.p1.s[j] = rand_in(-BRIEF_PATCH_SIZE_HALF, BRIEF_PATCH_SIZE_HALF + 1, &ctx->alfg);
1248 pair.p2.s[j] = rand_in(-BRIEF_PATCH_SIZE_HALF, BRIEF_PATCH_SIZE_HALF + 1, &ctx->alfg);
1251 pattern_host[i] = pair;
1254 for (int i = 0; i < 14; i++) {
1255 if (ctx->sw_format == disallowed_formats[i]) {
1256 av_log(avctx, AV_LOG_ERROR, "unsupported format in deshake_opencl.\n");
1257 err = AVERROR(ENOSYS);
1262 if (desc->flags & AV_PIX_FMT_FLAG_RGB) {
1263 ctx->is_yuv = false;
1267 ctx->sw_format = hw_frames_ctx->sw_format;
1269 err = ff_opencl_filter_load_program(avctx, &ff_opencl_source_deshake, 1);
1273 if (ctx->debug_on) {
1274 queue_props = CL_QUEUE_PROFILING_ENABLE;
1278 ctx->command_queue = clCreateCommandQueue(
1279 ctx->ocf.hwctx->context,
1280 ctx->ocf.hwctx->device_id,
1284 CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to create OpenCL command queue %d.\n", cle);
1286 CL_CREATE_KERNEL(ctx, grayscale);
1287 CL_CREATE_KERNEL(ctx, harris_response);
1288 CL_CREATE_KERNEL(ctx, refine_features);
1289 CL_CREATE_KERNEL(ctx, brief_descriptors);
1290 CL_CREATE_KERNEL(ctx, match_descriptors);
1291 CL_CREATE_KERNEL(ctx, transform);
1292 CL_CREATE_KERNEL(ctx, crop_upscale);
1294 CL_CREATE_KERNEL(ctx, draw_debug_info);
1297 grayscale_format.image_channel_order = CL_R;
1298 grayscale_format.image_channel_data_type = CL_FLOAT;
1300 grayscale_desc = (cl_image_desc) {
1301 .image_type = CL_MEM_OBJECT_IMAGE2D,
1302 .image_width = outlink->w,
1303 .image_height = outlink->h,
1305 .image_array_size = 0,
1306 .image_row_pitch = 0,
1307 .image_slice_pitch = 0,
1308 .num_mip_levels = 0,
1313 ctx->grayscale = clCreateImage(
1314 ctx->ocf.hwctx->context,
1321 CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to create grayscale image: %d.\n", cle);
1324 CL_CREATE_BUFFER(ctx, harris_buf, outlink->h * outlink->w * sizeof(float));
1325 CL_CREATE_BUFFER(ctx, refined_features, features_buf_size);
1326 CL_CREATE_BUFFER(ctx, prev_refined_features, features_buf_size);
1327 CL_CREATE_BUFFER_FLAGS(
1330 CL_MEM_READ_WRITE | CL_MEM_COPY_HOST_PTR,
1331 BREIFN * sizeof(PointPair),
1334 CL_CREATE_BUFFER(ctx, descriptors, descriptor_buf_size);
1335 CL_CREATE_BUFFER(ctx, prev_descriptors, descriptor_buf_size);
1336 CL_CREATE_BUFFER(ctx, matches, image_grid_32 * sizeof(MotionVector));
1337 CL_CREATE_BUFFER(ctx, matches_contig, MATCHES_CONTIG_SIZE * sizeof(MotionVector));
1338 CL_CREATE_BUFFER(ctx, transform_y, 9 * sizeof(float));
1339 CL_CREATE_BUFFER(ctx, transform_uv, 9 * sizeof(float));
1340 if (ctx->debug_on) {
1341 CL_CREATE_BUFFER(ctx, debug_matches, MATCHES_CONTIG_SIZE * sizeof(MotionVector));
1342 CL_CREATE_BUFFER(ctx, debug_model_matches, 3 * sizeof(MotionVector));
1345 ctx->initialized = 1;
1346 av_freep(&pattern_host);
1352 av_freep(&pattern_host);
1356 // Logs debug information about the transform data
1357 static void transform_debug(AVFilterContext *avctx, float *new_vals, float *old_vals, int curr_frame) {
1358 av_log(avctx, AV_LOG_VERBOSE,
1360 "\tframe moved from: %f x, %f y\n"
1361 "\t to: %f x, %f y\n"
1362 "\t rotated from: %f degrees\n"
1363 "\t to: %f degrees\n"
1364 "\t scaled from: %f x, %f y\n"
1365 "\t to: %f x, %f y\n"
1367 "\tframe moved by: %f x, %f y\n"
1368 "\t rotated by: %f degrees\n"
1369 "\t scaled by: %f x, %f y\n",
1371 old_vals[RingbufX], old_vals[RingbufY],
1372 new_vals[RingbufX], new_vals[RingbufY],
1373 old_vals[RingbufRot] * (180.0 / M_PI),
1374 new_vals[RingbufRot] * (180.0 / M_PI),
1375 old_vals[RingbufScaleX], old_vals[RingbufScaleY],
1376 new_vals[RingbufScaleX], new_vals[RingbufScaleY],
1377 old_vals[RingbufX] - new_vals[RingbufX], old_vals[RingbufY] - new_vals[RingbufY],
1378 old_vals[RingbufRot] * (180.0 / M_PI) - new_vals[RingbufRot] * (180.0 / M_PI),
1379 new_vals[RingbufScaleX] / old_vals[RingbufScaleX], new_vals[RingbufScaleY] / old_vals[RingbufScaleY]
1383 // Uses the buffered motion information to determine a transform that smooths the
1384 // given frame and applies it
1385 static int filter_frame(AVFilterLink *link, AVFrame *input_frame)
1387 AVFilterContext *avctx = link->dst;
1388 AVFilterLink *outlink = avctx->outputs[0];
1389 DeshakeOpenCLContext *deshake_ctx = avctx->priv;
1390 AVFrame *cropped_frame = NULL, *transformed_frame = NULL;
1393 float new_vals[RingbufCount];
1394 float old_vals[RingbufCount];
1395 // Luma (in the case of YUV) transform, or just the transform in the case of RGB
1396 float transform_y[9];
1398 float transform_uv[9];
1399 // Luma crop transform (or RGB)
1400 float transform_crop_y[9];
1401 // Chroma crop transform
1402 float transform_crop_uv[9];
1403 float transform_debug_rgb[9];
1404 size_t global_work[2];
1406 cl_mem src, transformed, dst;
1407 cl_mem transforms[3];
1409 cl_event transform_event, crop_upscale_event;
1410 DebugMatches debug_matches;
1411 cl_int num_model_matches;
1413 const float center_w = (float)input_frame->width / 2;
1414 const float center_h = (float)input_frame->height / 2;
1416 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(deshake_ctx->sw_format);
1417 const int chroma_width = AV_CEIL_RSHIFT(input_frame->width, desc->log2_chroma_w);
1418 const int chroma_height = AV_CEIL_RSHIFT(input_frame->height, desc->log2_chroma_h);
1420 const float center_w_chroma = (float)chroma_width / 2;
1421 const float center_h_chroma = (float)chroma_height / 2;
1423 const float luma_w_over_chroma_w = ((float)input_frame->width / (float)chroma_width);
1424 const float luma_h_over_chroma_h = ((float)input_frame->height / (float)chroma_height);
1426 if (deshake_ctx->debug_on) {
1427 av_fifo_generic_read(
1428 deshake_ctx->abs_motion.debug_matches,
1430 sizeof(DebugMatches),
1435 if (input_frame->pkt_duration) {
1436 duration = input_frame->pkt_duration;
1438 duration = av_rescale_q(1, av_inv_q(outlink->frame_rate), outlink->time_base);
1440 deshake_ctx->duration = input_frame->pts + duration;
1442 // Get the absolute transform data for this frame
1443 for (int i = 0; i < RingbufCount; i++) {
1444 av_fifo_generic_peek_at(
1445 deshake_ctx->abs_motion.ringbuffers[i],
1447 deshake_ctx->abs_motion.curr_frame_offset * sizeof(float),
1453 if (deshake_ctx->tripod_mode) {
1454 // If tripod mode is turned on we simply undo all motion relative to the
1457 new_vals[RingbufX] = 0.0f;
1458 new_vals[RingbufY] = 0.0f;
1459 new_vals[RingbufRot] = 0.0f;
1460 new_vals[RingbufScaleX] = 1.0f;
1461 new_vals[RingbufScaleY] = 1.0f;
1463 // Tripod mode is off and we need to smooth a moving camera
1465 new_vals[RingbufX] = smooth(
1467 deshake_ctx->gauss_kernel,
1468 deshake_ctx->smooth_window,
1470 deshake_ctx->abs_motion.ringbuffers[RingbufX]
1472 new_vals[RingbufY] = smooth(
1474 deshake_ctx->gauss_kernel,
1475 deshake_ctx->smooth_window,
1476 input_frame->height,
1477 deshake_ctx->abs_motion.ringbuffers[RingbufY]
1479 new_vals[RingbufRot] = smooth(
1481 deshake_ctx->gauss_kernel,
1482 deshake_ctx->smooth_window,
1484 deshake_ctx->abs_motion.ringbuffers[RingbufRot]
1486 new_vals[RingbufScaleX] = smooth(
1488 deshake_ctx->gauss_kernel,
1489 deshake_ctx->smooth_window,
1491 deshake_ctx->abs_motion.ringbuffers[RingbufScaleX]
1493 new_vals[RingbufScaleY] = smooth(
1495 deshake_ctx->gauss_kernel,
1496 deshake_ctx->smooth_window,
1498 deshake_ctx->abs_motion.ringbuffers[RingbufScaleY]
1502 transform_center_scale(
1503 old_vals[RingbufX] - new_vals[RingbufX],
1504 old_vals[RingbufY] - new_vals[RingbufY],
1505 old_vals[RingbufRot] - new_vals[RingbufRot],
1506 new_vals[RingbufScaleX] / old_vals[RingbufScaleX],
1507 new_vals[RingbufScaleY] / old_vals[RingbufScaleY],
1513 transform_center_scale(
1514 (old_vals[RingbufX] - new_vals[RingbufX]) / luma_w_over_chroma_w,
1515 (old_vals[RingbufY] - new_vals[RingbufY]) / luma_h_over_chroma_h,
1516 old_vals[RingbufRot] - new_vals[RingbufRot],
1517 new_vals[RingbufScaleX] / old_vals[RingbufScaleX],
1518 new_vals[RingbufScaleY] / old_vals[RingbufScaleY],
1524 CL_BLOCKING_WRITE_BUFFER(deshake_ctx->command_queue, deshake_ctx->transform_y, 9 * sizeof(float), transform_y, NULL);
1525 CL_BLOCKING_WRITE_BUFFER(deshake_ctx->command_queue, deshake_ctx->transform_uv, 9 * sizeof(float), transform_uv, NULL);
1527 if (deshake_ctx->debug_on)
1528 transform_debug(avctx, new_vals, old_vals, deshake_ctx->curr_frame);
1530 cropped_frame = ff_get_video_buffer(outlink, outlink->w, outlink->h);
1531 if (!cropped_frame) {
1532 err = AVERROR(ENOMEM);
1536 transformed_frame = ff_get_video_buffer(outlink, outlink->w, outlink->h);
1537 if (!transformed_frame) {
1538 err = AVERROR(ENOMEM);
1542 transforms[0] = deshake_ctx->transform_y;
1543 transforms[1] = transforms[2] = deshake_ctx->transform_uv;
1545 for (int p = 0; p < FF_ARRAY_ELEMS(transformed_frame->data); p++) {
1546 // Transform all of the planes appropriately
1547 src = (cl_mem)input_frame->data[p];
1548 transformed = (cl_mem)transformed_frame->data[p];
1553 err = ff_opencl_filter_work_size_from_image(avctx, global_work, input_frame, p, 0);
1557 CL_RUN_KERNEL_WITH_ARGS(
1558 deshake_ctx->command_queue,
1559 deshake_ctx->kernel_transform,
1563 { sizeof(cl_mem), &src },
1564 { sizeof(cl_mem), &transformed },
1565 { sizeof(cl_mem), &transforms[p] },
1569 if (deshake_ctx->debug_on && !deshake_ctx->is_yuv && debug_matches.num_matches > 0) {
1570 CL_BLOCKING_WRITE_BUFFER(
1571 deshake_ctx->command_queue,
1572 deshake_ctx->debug_matches,
1573 debug_matches.num_matches * sizeof(MotionVector),
1574 debug_matches.matches,
1578 CL_BLOCKING_WRITE_BUFFER(
1579 deshake_ctx->command_queue,
1580 deshake_ctx->debug_model_matches,
1581 debug_matches.num_model_matches * sizeof(MotionVector),
1582 debug_matches.model_matches,
1586 num_model_matches = debug_matches.num_model_matches;
1588 // Invert the transform
1589 transform_center_scale(
1590 new_vals[RingbufX] - old_vals[RingbufX],
1591 new_vals[RingbufY] - old_vals[RingbufY],
1592 new_vals[RingbufRot] - old_vals[RingbufRot],
1593 old_vals[RingbufScaleX] / new_vals[RingbufScaleX],
1594 old_vals[RingbufScaleY] / new_vals[RingbufScaleY],
1600 CL_BLOCKING_WRITE_BUFFER(deshake_ctx->command_queue, deshake_ctx->transform_y, 9 * sizeof(float), transform_debug_rgb, NULL);
1602 transformed = (cl_mem)transformed_frame->data[0];
1603 CL_RUN_KERNEL_WITH_ARGS(
1604 deshake_ctx->command_queue,
1605 deshake_ctx->kernel_draw_debug_info,
1606 (size_t[]){ debug_matches.num_matches },
1609 { sizeof(cl_mem), &transformed },
1610 { sizeof(cl_mem), &deshake_ctx->debug_matches },
1611 { sizeof(cl_mem), &deshake_ctx->debug_model_matches },
1612 { sizeof(cl_int), &num_model_matches },
1613 { sizeof(cl_mem), &deshake_ctx->transform_y }
1617 if (deshake_ctx->should_crop) {
1618 // Generate transforms for cropping
1619 transform_center_scale(
1620 (old_vals[RingbufX] - new_vals[RingbufX]) / 5,
1621 (old_vals[RingbufY] - new_vals[RingbufY]) / 5,
1622 (old_vals[RingbufRot] - new_vals[RingbufRot]) / 5,
1623 new_vals[RingbufScaleX] / old_vals[RingbufScaleX],
1624 new_vals[RingbufScaleY] / old_vals[RingbufScaleY],
1629 update_needed_crop(&deshake_ctx->crop_y, transform_crop_y, input_frame->width, input_frame->height);
1631 transform_center_scale(
1632 (old_vals[RingbufX] - new_vals[RingbufX]) / (5 * luma_w_over_chroma_w),
1633 (old_vals[RingbufY] - new_vals[RingbufY]) / (5 * luma_h_over_chroma_h),
1634 (old_vals[RingbufRot] - new_vals[RingbufRot]) / 5,
1635 new_vals[RingbufScaleX] / old_vals[RingbufScaleX],
1636 new_vals[RingbufScaleY] / old_vals[RingbufScaleY],
1641 update_needed_crop(&deshake_ctx->crop_uv, transform_crop_uv, chroma_width, chroma_height);
1643 crops[0] = deshake_ctx->crop_y;
1644 crops[1] = crops[2] = deshake_ctx->crop_uv;
1646 for (int p = 0; p < FF_ARRAY_ELEMS(cropped_frame->data); p++) {
1647 // Crop all of the planes appropriately
1648 dst = (cl_mem)cropped_frame->data[p];
1649 transformed = (cl_mem)transformed_frame->data[p];
1654 err = ff_opencl_filter_work_size_from_image(avctx, global_work, input_frame, p, 0);
1658 CL_RUN_KERNEL_WITH_ARGS(
1659 deshake_ctx->command_queue,
1660 deshake_ctx->kernel_crop_upscale,
1663 &crop_upscale_event,
1664 { sizeof(cl_mem), &transformed },
1665 { sizeof(cl_mem), &dst },
1666 { sizeof(cl_float2), &crops[p].top_left },
1667 { sizeof(cl_float2), &crops[p].bottom_right },
1672 if (deshake_ctx->curr_frame < deshake_ctx->smooth_window / 2) {
1673 // This means we are somewhere at the start of the video. We need to
1674 // increment the current frame offset until it reaches the center of
1675 // the ringbuffers (as the current frame will be located there for
1676 // the rest of the video).
1678 // The end of the video is taken care of by draining motion data
1679 // one-by-one out of the buffer, causing the (at that point fixed)
1680 // offset to move towards later frames' data.
1681 ++deshake_ctx->abs_motion.curr_frame_offset;
1684 if (deshake_ctx->abs_motion.data_end_offset != -1) {
1685 // Keep the end offset in sync with the frame it's supposed to be
1687 --deshake_ctx->abs_motion.data_end_offset;
1689 if (deshake_ctx->abs_motion.data_end_offset == deshake_ctx->abs_motion.curr_frame_offset - 1) {
1690 // The end offset would be the start of the new video sequence; flip to
1692 deshake_ctx->abs_motion.data_end_offset = -1;
1693 deshake_ctx->abs_motion.data_start_offset = deshake_ctx->abs_motion.curr_frame_offset;
1695 } else if (deshake_ctx->abs_motion.data_start_offset != -1) {
1696 // Keep the start offset in sync with the frame it's supposed to be
1698 --deshake_ctx->abs_motion.data_start_offset;
1701 if (deshake_ctx->debug_on) {
1702 deshake_ctx->transform_time += ff_opencl_get_event_time(transform_event);
1703 if (deshake_ctx->should_crop) {
1704 deshake_ctx->crop_upscale_time += ff_opencl_get_event_time(crop_upscale_event);
1708 ++deshake_ctx->curr_frame;
1710 if (deshake_ctx->debug_on)
1711 av_freep(&debug_matches.matches);
1713 if (deshake_ctx->should_crop) {
1714 err = av_frame_copy_props(cropped_frame, input_frame);
1718 av_frame_free(&transformed_frame);
1719 av_frame_free(&input_frame);
1720 return ff_filter_frame(outlink, cropped_frame);
1723 err = av_frame_copy_props(transformed_frame, input_frame);
1727 av_frame_free(&cropped_frame);
1728 av_frame_free(&input_frame);
1729 return ff_filter_frame(outlink, transformed_frame);
1733 clFinish(deshake_ctx->command_queue);
1735 if (deshake_ctx->debug_on)
1736 if (debug_matches.matches)
1737 av_freep(&debug_matches.matches);
1739 av_frame_free(&input_frame);
1740 av_frame_free(&transformed_frame);
1741 av_frame_free(&cropped_frame);
1745 // Add the given frame to the frame queue to eventually be processed.
1747 // Also determines the motion from the previous frame and updates the stored
1748 // motion information accordingly.
1749 static int queue_frame(AVFilterLink *link, AVFrame *input_frame)
1751 AVFilterContext *avctx = link->dst;
1752 DeshakeOpenCLContext *deshake_ctx = avctx->priv;
1755 int num_inliers = 0;
1757 FrameDelta relative;
1758 SimilarityMatrix model;
1759 size_t global_work[2];
1760 size_t harris_global_work[2];
1761 size_t grid_32_global_work[2];
1762 int grid_32_h, grid_32_w;
1763 size_t local_work[2];
1767 cl_event grayscale_event, harris_response_event, refine_features_event,
1768 brief_event, match_descriptors_event, read_buf_event;
1769 DebugMatches debug_matches;
1776 err = ff_opencl_filter_work_size_from_image(avctx, global_work, input_frame, 0, 0);
1780 err = ff_opencl_filter_work_size_from_image(avctx, harris_global_work, input_frame, 0, 8);
1784 err = ff_opencl_filter_work_size_from_image(avctx, grid_32_global_work, input_frame, 0, 32);
1788 // We want a single work-item for each 32x32 block of pixels in the input frame
1789 grid_32_global_work[0] /= 32;
1790 grid_32_global_work[1] /= 32;
1792 grid_32_h = ROUNDED_UP_DIV(input_frame->height, 32);
1793 grid_32_w = ROUNDED_UP_DIV(input_frame->width, 32);
1795 if (deshake_ctx->is_yuv) {
1796 deshake_ctx->grayscale = (cl_mem)input_frame->data[0];
1798 src = (cl_mem)input_frame->data[0];
1800 CL_RUN_KERNEL_WITH_ARGS(
1801 deshake_ctx->command_queue,
1802 deshake_ctx->kernel_grayscale,
1806 { sizeof(cl_mem), &src },
1807 { sizeof(cl_mem), &deshake_ctx->grayscale }
1811 CL_RUN_KERNEL_WITH_ARGS(
1812 deshake_ctx->command_queue,
1813 deshake_ctx->kernel_harris_response,
1816 &harris_response_event,
1817 { sizeof(cl_mem), &deshake_ctx->grayscale },
1818 { sizeof(cl_mem), &deshake_ctx->harris_buf }
1821 CL_RUN_KERNEL_WITH_ARGS(
1822 deshake_ctx->command_queue,
1823 deshake_ctx->kernel_refine_features,
1824 grid_32_global_work,
1826 &refine_features_event,
1827 { sizeof(cl_mem), &deshake_ctx->grayscale },
1828 { sizeof(cl_mem), &deshake_ctx->harris_buf },
1829 { sizeof(cl_mem), &deshake_ctx->refined_features },
1830 { sizeof(cl_int), &deshake_ctx->refine_features }
1833 CL_RUN_KERNEL_WITH_ARGS(
1834 deshake_ctx->command_queue,
1835 deshake_ctx->kernel_brief_descriptors,
1836 grid_32_global_work,
1839 { sizeof(cl_mem), &deshake_ctx->grayscale },
1840 { sizeof(cl_mem), &deshake_ctx->refined_features },
1841 { sizeof(cl_mem), &deshake_ctx->descriptors },
1842 { sizeof(cl_mem), &deshake_ctx->brief_pattern}
1845 if (av_fifo_size(deshake_ctx->abs_motion.ringbuffers[RingbufX]) == 0) {
1846 // This is the first frame we've been given to queue, meaning there is
1847 // no previous frame to match descriptors to
1849 goto no_motion_data;
1852 CL_RUN_KERNEL_WITH_ARGS(
1853 deshake_ctx->command_queue,
1854 deshake_ctx->kernel_match_descriptors,
1855 grid_32_global_work,
1857 &match_descriptors_event,
1858 { sizeof(cl_mem), &deshake_ctx->prev_refined_features },
1859 { sizeof(cl_mem), &deshake_ctx->refined_features },
1860 { sizeof(cl_mem), &deshake_ctx->descriptors },
1861 { sizeof(cl_mem), &deshake_ctx->prev_descriptors },
1862 { sizeof(cl_mem), &deshake_ctx->matches }
1865 cle = clEnqueueReadBuffer(
1866 deshake_ctx->command_queue,
1867 deshake_ctx->matches,
1870 grid_32_h * grid_32_w * sizeof(MotionVector),
1871 deshake_ctx->matches_host,
1876 CL_FAIL_ON_ERROR(AVERROR(EIO), "Failed to read matches to host: %d.\n", cle);
1878 num_vectors = make_vectors_contig(deshake_ctx, grid_32_h, grid_32_w);
1880 if (num_vectors < 10) {
1881 // Not enough matches to get reliable motion data for this frame
1883 // From this point on all data is relative to this frame rather than the
1884 // original frame. We have to make sure that we don't mix values that were
1885 // relative to the original frame with the new values relative to this
1886 // frame when doing the gaussian smoothing. We keep track of where the old
1887 // values end using this data_end_offset field in order to accomplish
1890 // If no motion data is present for multiple frames in a short window of
1891 // time, we leave the end where it was to avoid mixing 0s in with the
1892 // old data (and just treat them all as part of the new values)
1893 if (deshake_ctx->abs_motion.data_end_offset == -1) {
1894 deshake_ctx->abs_motion.data_end_offset =
1895 av_fifo_size(deshake_ctx->abs_motion.ringbuffers[RingbufX]) / sizeof(float) - 1;
1898 goto no_motion_data;
1901 if (!estimate_affine_2d(
1903 deshake_ctx->matches_contig_host,
1911 goto no_motion_data;
1914 for (int i = 0; i < num_vectors; i++) {
1915 if (deshake_ctx->matches_contig_host[i].should_consider) {
1916 deshake_ctx->inliers[num_inliers] = deshake_ctx->matches_contig_host[i];
1921 if (!minimize_error(
1923 deshake_ctx->inliers,
1929 goto no_motion_data;
1933 relative = decompose_transform(model.matrix);
1935 // Get the absolute transform data for the previous frame
1936 for (int i = 0; i < RingbufCount; i++) {
1937 av_fifo_generic_peek_at(
1938 deshake_ctx->abs_motion.ringbuffers[i],
1940 av_fifo_size(deshake_ctx->abs_motion.ringbuffers[i]) - sizeof(float),
1946 new_vals[RingbufX] = prev_vals[RingbufX] + relative.translation.s[0];
1947 new_vals[RingbufY] = prev_vals[RingbufY] + relative.translation.s[1];
1948 new_vals[RingbufRot] = prev_vals[RingbufRot] + relative.rotation;
1949 new_vals[RingbufScaleX] = prev_vals[RingbufScaleX] / relative.scale.s[0];
1950 new_vals[RingbufScaleY] = prev_vals[RingbufScaleY] / relative.scale.s[1];
1952 if (deshake_ctx->debug_on) {
1953 if (!deshake_ctx->is_yuv) {
1954 deshake_ctx->grayscale_time += ff_opencl_get_event_time(grayscale_event);
1956 deshake_ctx->harris_response_time += ff_opencl_get_event_time(harris_response_event);
1957 deshake_ctx->refine_features_time += ff_opencl_get_event_time(refine_features_event);
1958 deshake_ctx->brief_descriptors_time += ff_opencl_get_event_time(brief_event);
1959 deshake_ctx->match_descriptors_time += ff_opencl_get_event_time(match_descriptors_event);
1960 deshake_ctx->read_buf_time += ff_opencl_get_event_time(read_buf_event);
1966 new_vals[RingbufX] = 0.0f;
1967 new_vals[RingbufY] = 0.0f;
1968 new_vals[RingbufRot] = 0.0f;
1969 new_vals[RingbufScaleX] = 1.0f;
1970 new_vals[RingbufScaleY] = 1.0f;
1972 for (int i = 0; i < num_vectors; i++) {
1973 deshake_ctx->matches_contig_host[i].should_consider = false;
1975 debug_matches.num_model_matches = 0;
1977 if (deshake_ctx->debug_on) {
1978 av_log(avctx, AV_LOG_VERBOSE,
1979 "\n[ALERT] No motion data found in queue_frame, motion reset to 0\n\n"
1986 // Swap the descriptor buffers (we don't need the previous frame's descriptors
1987 // again so we will use that space for the next frame's descriptors)
1988 temp = deshake_ctx->prev_descriptors;
1989 deshake_ctx->prev_descriptors = deshake_ctx->descriptors;
1990 deshake_ctx->descriptors = temp;
1992 // Same for the refined features
1993 temp = deshake_ctx->prev_refined_features;
1994 deshake_ctx->prev_refined_features = deshake_ctx->refined_features;
1995 deshake_ctx->refined_features = temp;
1997 if (deshake_ctx->debug_on) {
1998 if (num_vectors == 0) {
1999 debug_matches.matches = NULL;
2001 debug_matches.matches = av_malloc_array(num_vectors, sizeof(MotionVector));
2003 if (!debug_matches.matches) {
2004 err = AVERROR(ENOMEM);
2009 for (int i = 0; i < num_vectors; i++) {
2010 debug_matches.matches[i] = deshake_ctx->matches_contig_host[i];
2012 debug_matches.num_matches = num_vectors;
2014 av_fifo_generic_write(
2015 deshake_ctx->abs_motion.debug_matches,
2017 sizeof(DebugMatches),
2022 for (int i = 0; i < RingbufCount; i++) {
2023 av_fifo_generic_write(
2024 deshake_ctx->abs_motion.ringbuffers[i],
2031 return ff_framequeue_add(&deshake_ctx->fq, input_frame);
2034 clFinish(deshake_ctx->command_queue);
2035 av_frame_free(&input_frame);
2039 static int activate(AVFilterContext *ctx)
2041 AVFilterLink *inlink = ctx->inputs[0];
2042 AVFilterLink *outlink = ctx->outputs[0];
2043 DeshakeOpenCLContext *deshake_ctx = ctx->priv;
2044 AVFrame *frame = NULL;
2048 FF_FILTER_FORWARD_STATUS_BACK(outlink, inlink);
2050 if (!deshake_ctx->eof) {
2051 ret = ff_inlink_consume_frame(inlink, &frame);
2055 if (!frame->hw_frames_ctx)
2056 return AVERROR(EINVAL);
2058 if (!deshake_ctx->initialized) {
2059 ret = deshake_opencl_init(ctx);
2064 // If there is no more space in the ringbuffers, remove the oldest
2065 // values to make room for the new ones
2066 if (av_fifo_space(deshake_ctx->abs_motion.ringbuffers[RingbufX]) == 0) {
2067 for (int i = 0; i < RingbufCount; i++) {
2068 av_fifo_drain(deshake_ctx->abs_motion.ringbuffers[i], sizeof(float));
2071 ret = queue_frame(inlink, frame);
2075 // See if we have enough buffered frames to process one
2077 // "enough" is half the smooth window of queued frames into the future
2078 if (ff_framequeue_queued_frames(&deshake_ctx->fq) >= deshake_ctx->smooth_window / 2) {
2079 return filter_frame(inlink, ff_framequeue_take(&deshake_ctx->fq));
2085 if (!deshake_ctx->eof && ff_inlink_acknowledge_status(inlink, &status, &pts)) {
2086 if (status == AVERROR_EOF) {
2087 deshake_ctx->eof = true;
2091 if (deshake_ctx->eof) {
2092 // Finish processing the rest of the frames in the queue.
2093 while(ff_framequeue_queued_frames(&deshake_ctx->fq) != 0) {
2094 for (int i = 0; i < RingbufCount; i++) {
2095 av_fifo_drain(deshake_ctx->abs_motion.ringbuffers[i], sizeof(float));
2098 ret = filter_frame(inlink, ff_framequeue_take(&deshake_ctx->fq));
2104 if (deshake_ctx->debug_on) {
2105 av_log(ctx, AV_LOG_VERBOSE,
2106 "Average kernel execution times:\n"
2107 "\t grayscale: %0.3f ms\n"
2108 "\t harris_response: %0.3f ms\n"
2109 "\t refine_features: %0.3f ms\n"
2110 "\tbrief_descriptors: %0.3f ms\n"
2111 "\tmatch_descriptors: %0.3f ms\n"
2112 "\t transform: %0.3f ms\n"
2113 "\t crop_upscale: %0.3f ms\n"
2114 "Average buffer read times:\n"
2115 "\t features buf: %0.3f ms\n",
2116 averaged_event_time_ms(deshake_ctx->grayscale_time, deshake_ctx->curr_frame),
2117 averaged_event_time_ms(deshake_ctx->harris_response_time, deshake_ctx->curr_frame),
2118 averaged_event_time_ms(deshake_ctx->refine_features_time, deshake_ctx->curr_frame),
2119 averaged_event_time_ms(deshake_ctx->brief_descriptors_time, deshake_ctx->curr_frame),
2120 averaged_event_time_ms(deshake_ctx->match_descriptors_time, deshake_ctx->curr_frame),
2121 averaged_event_time_ms(deshake_ctx->transform_time, deshake_ctx->curr_frame),
2122 averaged_event_time_ms(deshake_ctx->crop_upscale_time, deshake_ctx->curr_frame),
2123 averaged_event_time_ms(deshake_ctx->read_buf_time, deshake_ctx->curr_frame)
2127 ff_outlink_set_status(outlink, AVERROR_EOF, deshake_ctx->duration);
2131 if (!deshake_ctx->eof) {
2132 FF_FILTER_FORWARD_WANTED(outlink, inlink);
2135 return FFERROR_NOT_READY;
2138 static const AVFilterPad deshake_opencl_inputs[] = {
2141 .type = AVMEDIA_TYPE_VIDEO,
2142 .config_props = &ff_opencl_filter_config_input,
2147 static const AVFilterPad deshake_opencl_outputs[] = {
2150 .type = AVMEDIA_TYPE_VIDEO,
2151 .config_props = &ff_opencl_filter_config_output,
2156 #define OFFSET(x) offsetof(DeshakeOpenCLContext, x)
2157 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
2159 static const AVOption deshake_opencl_options[] = {
2161 "tripod", "simulates a tripod by preventing any camera movement whatsoever "
2162 "from the original frame",
2163 OFFSET(tripod_mode), AV_OPT_TYPE_BOOL, {.i64 = 0}, 0, 1, FLAGS
2166 "debug", "turn on additional debugging information",
2167 OFFSET(debug_on), AV_OPT_TYPE_BOOL, {.i64 = 0}, 0, 1, FLAGS
2170 "adaptive_crop", "attempt to subtly crop borders to reduce mirrored content",
2171 OFFSET(should_crop), AV_OPT_TYPE_BOOL, {.i64 = 1}, 0, 1, FLAGS
2174 "refine_features", "refine feature point locations at a sub-pixel level",
2175 OFFSET(refine_features), AV_OPT_TYPE_BOOL, {.i64 = 1}, 0, 1, FLAGS
2178 "smooth_strength", "smoothing strength (0 attempts to adaptively determine optimal strength)",
2179 OFFSET(smooth_percent), AV_OPT_TYPE_FLOAT, {.dbl = 0.0f}, 0.0f, 1.0f, FLAGS
2182 "smooth_window_multiplier", "multiplier for number of frames to buffer for motion data",
2183 OFFSET(smooth_window_multiplier), AV_OPT_TYPE_FLOAT, {.dbl = 2.0}, 0.1, 10.0, FLAGS
2188 AVFILTER_DEFINE_CLASS(deshake_opencl);
2190 AVFilter ff_vf_deshake_opencl = {
2191 .name = "deshake_opencl",
2192 .description = NULL_IF_CONFIG_SMALL("Feature-point based video stabilization filter"),
2193 .priv_size = sizeof(DeshakeOpenCLContext),
2194 .priv_class = &deshake_opencl_class,
2195 .init = &ff_opencl_filter_init,
2196 .uninit = &deshake_opencl_uninit,
2197 .query_formats = &ff_opencl_filter_query_formats,
2198 .activate = activate,
2199 .inputs = deshake_opencl_inputs,
2200 .outputs = deshake_opencl_outputs,
2201 .flags_internal = FF_FILTER_FLAG_HWFRAME_AWARE