2 * Copyright (C) 2010 Georg Martius <georg.martius@web.de>
3 * Copyright (C) 2010 Daniel G. Taylor <dan@programmer-art.org>
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * fast deshake / depan video filter
26 * SAD block-matching motion compensation to fix small changes in
27 * horizontal and/or vertical shift. This filter helps remove camera shake
28 * from hand-holding a camera, bumping a tripod, moving on a vehicle, etc.
31 * - For each frame with one previous reference frame
32 * - For each block in the frame
33 * - If contrast > threshold then find likely motion vector
34 * - For all found motion vectors
35 * - Find most common, store as global motion vector
36 * - Find most likely rotation angle
37 * - Transform image along global motion
40 * - Fill frame edges based on previous/next reference frames
41 * - Fill frame edges by stretching image near the edges?
42 * - Can this be done quickly and look decent?
44 * Dark Shikari links to http://wiki.videolan.org/SoC_x264_2010#GPU_Motion_Estimation_2
45 * for an algorithm similar to what could be used here to get the gmv
46 * It requires only a couple diamond searches + fast downscaling
48 * Special thanks to Jason Kotenko for his help with the algorithm and my
49 * inability to see simple errors in C code.
55 #include "libavutil/common.h"
56 #include "libavutil/mem.h"
57 #include "libavutil/pixdesc.h"
58 #include "libavcodec/dsputil.h"
60 #include "transform.h"
62 #define CHROMA_WIDTH(link) -((-link->w) >> av_pix_fmt_descriptors[link->format].log2_chroma_w)
63 #define CHROMA_HEIGHT(link) -((-link->h) >> av_pix_fmt_descriptors[link->format].log2_chroma_h)
66 EXHAUSTIVE, ///< Search all possible positions
67 SMART_EXHAUSTIVE, ///< Search most possible positions (faster)
72 int x; ///< Horizontal shift
73 int y; ///< Vertical shift
77 double x; ///< Horizontal shift
78 double y; ///< Vertical shift
82 MotionVector vector; ///< Motion vector
83 double angle; ///< Angle of rotation
84 double zoom; ///< Zoom percentage
89 AVFilterBufferRef *ref; ///< Previous frame
90 int rx; ///< Maximum horizontal shift
91 int ry; ///< Maximum vertical shift
92 enum FillMethod edge; ///< Edge fill method
93 int blocksize; ///< Size of blocks to compare
94 int contrast; ///< Contrast threshold
95 enum SearchMethod search; ///< Motion search method
96 AVCodecContext *avctx;
97 DSPContext c; ///< Context providing optimized SAD methods
98 Transform last; ///< Transform from last frame
99 int refcount; ///< Number of reference frames (defines averaging window)
102 int cw; ///< Crop motion search to this box
108 static int cmp(const double *a, const double *b)
110 return *a < *b ? -1 : ( *a > *b ? 1 : 0 );
114 * Cleaned mean (cuts off 20% of values to remove outliers and then averages)
116 static double clean_mean(double *values, int count)
122 qsort(values, count, sizeof(double), (void*)cmp);
124 for (x = cut; x < count - cut; x++) {
128 return mean / (count - cut * 2);
132 * Find the most likely shift in motion between two frames for a given
133 * macroblock. Test each block against several shifts given by the rx
134 * and ry attributes. Searches using a simple matrix of those shifts and
135 * chooses the most likely shift by the smallest difference in blocks.
137 static void find_block_motion(DeshakeContext *deshake, uint8_t *src1,
138 uint8_t *src2, int cx, int cy, int stride,
143 int smallest = INT_MAX;
146 #define CMP(i, j) deshake->c.sad[0](deshake, src1 + cy * stride + cx, \
147 src2 + (j) * stride + (i), stride, \
150 if (deshake->search == EXHAUSTIVE) {
151 // Compare every possible position - this is sloooow!
152 for (y = -deshake->ry; y <= deshake->ry; y++) {
153 for (x = -deshake->rx; x <= deshake->rx; x++) {
154 diff = CMP(cx - x, cy - y);
155 if (diff < smallest) {
162 } else if (deshake->search == SMART_EXHAUSTIVE) {
163 // Compare every other possible position and find the best match
164 for (y = -deshake->ry + 1; y < deshake->ry - 2; y += 2) {
165 for (x = -deshake->rx + 1; x < deshake->rx - 2; x += 2) {
166 diff = CMP(cx - x, cy - y);
167 if (diff < smallest) {
175 // Hone in on the specific best match around the match we found above
179 for (y = tmp2 - 1; y <= tmp2 + 1; y++) {
180 for (x = tmp - 1; x <= tmp + 1; x++) {
181 if (x == tmp && y == tmp2)
184 diff = CMP(cx - x, cy - y);
185 if (diff < smallest) {
194 if (smallest > 512) {
199 //av_log(NULL, AV_LOG_ERROR, "%d\n", smallest);
200 //av_log(NULL, AV_LOG_ERROR, "Final: (%d, %d) = %d x %d\n", cx, cy, mv->x, mv->y);
204 * Find the contrast of a given block. When searching for global motion we
205 * really only care about the high contrast blocks, so using this method we
206 * can actually skip blocks we don't care much about.
208 static int block_contrast(uint8_t *src, int x, int y, int stride, int blocksize)
214 for (i = 0; i <= blocksize * 2; i++) {
215 // We use a width of 16 here to match the libavcodec sad functions
216 for (j = 0; i <= 15; i++) {
217 pos = (y - i) * stride + (x - j);
218 if (src[pos] < lowest)
220 else if (src[pos] > highest) {
226 return highest - lowest;
230 * Find the rotation for a given block.
232 static double block_angle(int x, int y, int cx, int cy, IntMotionVector *shift)
236 a1 = atan2(y - cy, x - cx);
237 a2 = atan2(y - cy + shift->y, x - cx + shift->x);
241 return (diff > M_PI) ? diff - 2 * M_PI :
242 (diff < -M_PI) ? diff + 2 * M_PI :
247 * Find the estimated global motion for a scene given the most likely shift
248 * for each block in the frame. The global motion is estimated to be the
249 * same as the motion from most blocks in the frame, so if most blocks
250 * move one pixel to the right and two pixels down, this would yield a
251 * motion vector (1, -2).
253 static void find_motion(DeshakeContext *deshake, uint8_t *src1, uint8_t *src2,
254 int width, int height, int stride, Transform *t)
257 IntMotionVector mv = {0, 0};
258 int counts[128][128];
259 int count_max_value = 0;
263 double *angles = av_malloc(sizeof(*angles) * width * height / (16 * deshake->blocksize));
264 int center_x = 0, center_y = 0;
267 // Reset counts to zero
268 for (x = 0; x < deshake->rx * 2 + 1; x++) {
269 for (y = 0; y < deshake->ry * 2 + 1; y++) {
275 // Find motion for every block and store the motion vector in the counts
276 for (y = deshake->ry; y < height - deshake->ry - (deshake->blocksize * 2); y += deshake->blocksize * 2) {
277 // We use a width of 16 here to match the libavcodec sad functions
278 for (x = deshake->rx; x < width - deshake->rx - 16; x += 16) {
279 // If the contrast is too low, just skip this block as it probably
280 // won't be very useful to us.
281 contrast = block_contrast(src2, x, y, stride, deshake->blocksize);
282 if (contrast > deshake->contrast) {
283 //av_log(NULL, AV_LOG_ERROR, "%d\n", contrast);
284 find_block_motion(deshake, src1, src2, x, y, stride, &mv);
285 if (mv.x != -1 && mv.y != -1) {
286 counts[mv.x + deshake->rx][mv.y + deshake->ry] += 1;
287 if (x > deshake->rx && y > deshake->ry)
288 angles[pos++] = block_angle(x, y, 0, 0, &mv);
300 t->angle = clean_mean(angles, pos);
301 if (t->angle < 0.001)
307 // Find the most common motion vector in the frame and use it as the gmv
308 for (y = deshake->ry * 2; y >= 0; y--) {
309 for (x = 0; x < deshake->rx * 2 + 1; x++) {
310 //av_log(NULL, AV_LOG_ERROR, "%5d ", counts[x][y]);
311 if (counts[x][y] > count_max_value) {
312 t->vector.x = x - deshake->rx;
313 t->vector.y = y - deshake->ry;
314 count_max_value = counts[x][y];
317 //av_log(NULL, AV_LOG_ERROR, "\n");
320 p_x = (center_x - width / 2);
321 p_y = (center_y - height / 2);
322 t->vector.x += (cos(t->angle)-1)*p_x - sin(t->angle)*p_y;
323 t->vector.y += sin(t->angle)*p_x + (cos(t->angle)-1)*p_y;
325 // Clamp max shift & rotation?
326 t->vector.x = av_clipf(t->vector.x, -deshake->rx * 2, deshake->rx * 2);
327 t->vector.y = av_clipf(t->vector.y, -deshake->ry * 2, deshake->ry * 2);
328 t->angle = av_clipf(t->angle, -0.1, 0.1);
330 //av_log(NULL, AV_LOG_ERROR, "%d x %d\n", avg->x, avg->y);
334 static av_cold int init(AVFilterContext *ctx, const char *args)
336 DeshakeContext *deshake = ctx->priv;
337 char filename[256] = {0};
341 deshake->edge = FILL_MIRROR;
342 deshake->blocksize = 8;
343 deshake->contrast = 125;
344 deshake->search = EXHAUSTIVE;
345 deshake->refcount = 20;
353 sscanf(args, "%d:%d:%d:%d:%d:%d:%d:%d:%d:%d:%255s",
354 &deshake->cx, &deshake->cy, &deshake->cw, &deshake->ch,
355 &deshake->rx, &deshake->ry, (int *)&deshake->edge,
356 &deshake->blocksize, &deshake->contrast, (int *)&deshake->search, filename);
358 deshake->blocksize /= 2;
360 deshake->rx = av_clip(deshake->rx, 0, 64);
361 deshake->ry = av_clip(deshake->ry, 0, 64);
362 deshake->edge = av_clip(deshake->edge, FILL_BLANK, FILL_COUNT - 1);
363 deshake->blocksize = av_clip(deshake->blocksize, 4, 128);
364 deshake->contrast = av_clip(deshake->contrast, 1, 255);
365 deshake->search = av_clip(deshake->search, EXHAUSTIVE, SEARCH_COUNT - 1);
369 deshake->fp = fopen(filename, "w");
371 fwrite("Ori x, Avg x, Fin x, Ori y, Avg y, Fin y, Ori angle, Avg angle, Fin angle, Ori zoom, Avg zoom, Fin zoom\n", sizeof(char), 104, deshake->fp);
373 // Quadword align left edge of box for MMX code, adjust width if necessary
374 // to keep right margin
375 if (deshake->cx > 0) {
376 deshake->cw += deshake->cx - (deshake->cx & ~15);
380 av_log(ctx, AV_LOG_VERBOSE, "cx: %d, cy: %d, cw: %d, ch: %d, rx: %d, ry: %d, edge: %d blocksize: %d contrast: %d search: %d\n",
381 deshake->cx, deshake->cy, deshake->cw, deshake->ch,
382 deshake->rx, deshake->ry, deshake->edge, deshake->blocksize * 2, deshake->contrast, deshake->search);
387 static int query_formats(AVFilterContext *ctx)
389 enum AVPixelFormat pix_fmts[] = {
390 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV410P,
391 AV_PIX_FMT_YUV411P, AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ422P,
392 AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ440P, AV_PIX_FMT_NONE
395 ff_set_common_formats(ctx, ff_make_format_list(pix_fmts));
400 static int config_props(AVFilterLink *link)
402 DeshakeContext *deshake = link->dst->priv;
405 deshake->last.vector.x = 0;
406 deshake->last.vector.y = 0;
407 deshake->last.angle = 0;
408 deshake->last.zoom = 0;
410 deshake->avctx = avcodec_alloc_context3(NULL);
411 dsputil_init(&deshake->c, deshake->avctx);
416 static av_cold void uninit(AVFilterContext *ctx)
418 DeshakeContext *deshake = ctx->priv;
420 avfilter_unref_buffer(deshake->ref);
424 avcodec_close(deshake->avctx);
425 av_freep(&deshake->avctx);
428 static int end_frame(AVFilterLink *link)
430 DeshakeContext *deshake = link->dst->priv;
431 AVFilterBufferRef *in = link->cur_buf;
432 AVFilterBufferRef *out = link->dst->outputs[0]->out_buf;
433 Transform t = {{0},0}, orig = {{0},0};
435 float alpha = 2.0 / deshake->refcount;
438 link->cur_buf = NULL; /* it is in 'in' now */
439 if (deshake->cx < 0 || deshake->cy < 0 || deshake->cw < 0 || deshake->ch < 0) {
440 // Find the most likely global motion for the current frame
441 find_motion(deshake, (deshake->ref == NULL) ? in->data[0] : deshake->ref->data[0], in->data[0], link->w, link->h, in->linesize[0], &t);
443 uint8_t *src1 = (deshake->ref == NULL) ? in->data[0] : deshake->ref->data[0];
444 uint8_t *src2 = in->data[0];
446 deshake->cx = FFMIN(deshake->cx, link->w);
447 deshake->cy = FFMIN(deshake->cy, link->h);
449 if ((unsigned)deshake->cx + (unsigned)deshake->cw > link->w) deshake->cw = link->w - deshake->cx;
450 if ((unsigned)deshake->cy + (unsigned)deshake->ch > link->h) deshake->ch = link->h - deshake->cy;
452 // Quadword align right margin
455 src1 += deshake->cy * in->linesize[0] + deshake->cx;
456 src2 += deshake->cy * in->linesize[0] + deshake->cx;
458 find_motion(deshake, src1, src2, deshake->cw, deshake->ch, in->linesize[0], &t);
462 // Copy transform so we can output it later to compare to the smoothed value
463 orig.vector.x = t.vector.x;
464 orig.vector.y = t.vector.y;
465 orig.angle = t.angle;
468 // Generate a one-sided moving exponential average
469 deshake->avg.vector.x = alpha * t.vector.x + (1.0 - alpha) * deshake->avg.vector.x;
470 deshake->avg.vector.y = alpha * t.vector.y + (1.0 - alpha) * deshake->avg.vector.y;
471 deshake->avg.angle = alpha * t.angle + (1.0 - alpha) * deshake->avg.angle;
472 deshake->avg.zoom = alpha * t.zoom + (1.0 - alpha) * deshake->avg.zoom;
474 // Remove the average from the current motion to detect the motion that
475 // is not on purpose, just as jitter from bumping the camera
476 t.vector.x -= deshake->avg.vector.x;
477 t.vector.y -= deshake->avg.vector.y;
478 t.angle -= deshake->avg.angle;
479 t.zoom -= deshake->avg.zoom;
481 // Invert the motion to undo it
486 // Write statistics to file
488 snprintf(tmp, 256, "%f, %f, %f, %f, %f, %f, %f, %f, %f, %f, %f, %f\n", orig.vector.x, deshake->avg.vector.x, t.vector.x, orig.vector.y, deshake->avg.vector.y, t.vector.y, orig.angle, deshake->avg.angle, t.angle, orig.zoom, deshake->avg.zoom, t.zoom);
489 fwrite(tmp, sizeof(char), strlen(tmp), deshake->fp);
492 // Turn relative current frame motion into absolute by adding it to the
493 // last absolute motion
494 t.vector.x += deshake->last.vector.x;
495 t.vector.y += deshake->last.vector.y;
496 t.angle += deshake->last.angle;
497 t.zoom += deshake->last.zoom;
499 // Shrink motion by 10% to keep things centered in the camera frame
504 // Store the last absolute motion information
505 deshake->last.vector.x = t.vector.x;
506 deshake->last.vector.y = t.vector.y;
507 deshake->last.angle = t.angle;
508 deshake->last.zoom = t.zoom;
510 // Generate a luma transformation matrix
511 avfilter_get_matrix(t.vector.x, t.vector.y, t.angle, 1.0 + t.zoom / 100.0, matrix);
513 // Transform the luma plane
514 avfilter_transform(in->data[0], out->data[0], in->linesize[0], out->linesize[0], link->w, link->h, matrix, INTERPOLATE_BILINEAR, deshake->edge);
516 // Generate a chroma transformation matrix
517 avfilter_get_matrix(t.vector.x / (link->w / CHROMA_WIDTH(link)), t.vector.y / (link->h / CHROMA_HEIGHT(link)), t.angle, 1.0 + t.zoom / 100.0, matrix);
519 // Transform the chroma planes
520 avfilter_transform(in->data[1], out->data[1], in->linesize[1], out->linesize[1], CHROMA_WIDTH(link), CHROMA_HEIGHT(link), matrix, INTERPOLATE_BILINEAR, deshake->edge);
521 avfilter_transform(in->data[2], out->data[2], in->linesize[2], out->linesize[2], CHROMA_WIDTH(link), CHROMA_HEIGHT(link), matrix, INTERPOLATE_BILINEAR, deshake->edge);
523 // Store the current frame as the reference frame for calculating the
524 // motion of the next frame
525 if (deshake->ref != NULL)
526 avfilter_unref_buffer(deshake->ref);
528 // Cleanup the old reference frame
531 // Draw the transformed frame information
532 ff_draw_slice(link->dst->outputs[0], 0, link->h, 1);
533 return ff_end_frame(link->dst->outputs[0]);
536 static int draw_slice(AVFilterLink *link, int y, int h, int slice_dir)
541 AVFilter avfilter_vf_deshake = {
543 .description = NULL_IF_CONFIG_SMALL("Stabilize shaky video."),
545 .priv_size = sizeof(DeshakeContext),
549 .query_formats = query_formats,
551 .inputs = (const AVFilterPad[]) {{ .name = "default",
552 .type = AVMEDIA_TYPE_VIDEO,
553 .draw_slice = draw_slice,
554 .end_frame = end_frame,
555 .config_props = config_props,
556 .min_perms = AV_PERM_READ | AV_PERM_PRESERVE, },
559 .outputs = (const AVFilterPad[]) {{ .name = "default",
560 .type = AVMEDIA_TYPE_VIDEO, },