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.
56 #include "libavutil/common.h"
57 #include "libavutil/mem.h"
58 #include "libavutil/opt.h"
59 #include "libavutil/pixdesc.h"
60 #include "libavutil/qsort.h"
64 #define OFFSET(x) offsetof(DeshakeContext, x)
65 #define FLAGS AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
67 static const AVOption deshake_options[] = {
68 { "x", "set x for the rectangular search area", OFFSET(cx), AV_OPT_TYPE_INT, {.i64=-1}, -1, INT_MAX, .flags = FLAGS },
69 { "y", "set y for the rectangular search area", OFFSET(cy), AV_OPT_TYPE_INT, {.i64=-1}, -1, INT_MAX, .flags = FLAGS },
70 { "w", "set width for the rectangular search area", OFFSET(cw), AV_OPT_TYPE_INT, {.i64=-1}, -1, INT_MAX, .flags = FLAGS },
71 { "h", "set height for the rectangular search area", OFFSET(ch), AV_OPT_TYPE_INT, {.i64=-1}, -1, INT_MAX, .flags = FLAGS },
72 { "rx", "set x for the rectangular search area", OFFSET(rx), AV_OPT_TYPE_INT, {.i64=16}, 0, MAX_R, .flags = FLAGS },
73 { "ry", "set y for the rectangular search area", OFFSET(ry), AV_OPT_TYPE_INT, {.i64=16}, 0, MAX_R, .flags = FLAGS },
74 { "edge", "set edge mode", OFFSET(edge), AV_OPT_TYPE_INT, {.i64=FILL_MIRROR}, FILL_BLANK, FILL_COUNT-1, FLAGS, "edge"},
75 { "blank", "fill zeroes at blank locations", 0, AV_OPT_TYPE_CONST, {.i64=FILL_BLANK}, INT_MIN, INT_MAX, FLAGS, "edge" },
76 { "original", "original image at blank locations", 0, AV_OPT_TYPE_CONST, {.i64=FILL_ORIGINAL}, INT_MIN, INT_MAX, FLAGS, "edge" },
77 { "clamp", "extruded edge value at blank locations", 0, AV_OPT_TYPE_CONST, {.i64=FILL_CLAMP}, INT_MIN, INT_MAX, FLAGS, "edge" },
78 { "mirror", "mirrored edge at blank locations", 0, AV_OPT_TYPE_CONST, {.i64=FILL_MIRROR}, INT_MIN, INT_MAX, FLAGS, "edge" },
79 { "blocksize", "set motion search blocksize", OFFSET(blocksize), AV_OPT_TYPE_INT, {.i64=8}, 4, 128, .flags = FLAGS },
80 { "contrast", "set contrast threshold for blocks", OFFSET(contrast), AV_OPT_TYPE_INT, {.i64=125}, 1, 255, .flags = FLAGS },
81 { "search", "set search strategy", OFFSET(search), AV_OPT_TYPE_INT, {.i64=EXHAUSTIVE}, EXHAUSTIVE, SEARCH_COUNT-1, FLAGS, "smode" },
82 { "exhaustive", "exhaustive search", 0, AV_OPT_TYPE_CONST, {.i64=EXHAUSTIVE}, INT_MIN, INT_MAX, FLAGS, "smode" },
83 { "less", "less exhaustive search", 0, AV_OPT_TYPE_CONST, {.i64=SMART_EXHAUSTIVE}, INT_MIN, INT_MAX, FLAGS, "smode" },
84 { "filename", "set motion search detailed log file name", OFFSET(filename), AV_OPT_TYPE_STRING, {.str=NULL}, .flags = FLAGS },
85 { "opencl", "ignored", OFFSET(opencl), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, .flags = FLAGS },
89 AVFILTER_DEFINE_CLASS(deshake);
91 static int cmp(const void *a, const void *b)
93 return FFDIFFSIGN(*(const double *)a, *(const double *)b);
97 * Cleaned mean (cuts off 20% of values to remove outliers and then averages)
99 static double clean_mean(double *values, int count)
105 AV_QSORT(values, count, double, cmp);
107 for (x = cut; x < count - cut; x++) {
111 return mean / (count - cut * 2);
115 * Find the most likely shift in motion between two frames for a given
116 * macroblock. Test each block against several shifts given by the rx
117 * and ry attributes. Searches using a simple matrix of those shifts and
118 * chooses the most likely shift by the smallest difference in blocks.
120 static void find_block_motion(DeshakeContext *deshake, uint8_t *src1,
121 uint8_t *src2, int cx, int cy, int stride,
126 int smallest = INT_MAX;
129 #define CMP(i, j) deshake->sad(src1 + cy * stride + cx, stride,\
130 src2 + (j) * stride + (i), stride)
132 if (deshake->search == EXHAUSTIVE) {
133 // Compare every possible position - this is sloooow!
134 for (y = -deshake->ry; y <= deshake->ry; y++) {
135 for (x = -deshake->rx; x <= deshake->rx; x++) {
136 diff = CMP(cx - x, cy - y);
137 if (diff < smallest) {
144 } else if (deshake->search == SMART_EXHAUSTIVE) {
145 // Compare every other possible position and find the best match
146 for (y = -deshake->ry + 1; y < deshake->ry; y += 2) {
147 for (x = -deshake->rx + 1; x < deshake->rx; x += 2) {
148 diff = CMP(cx - x, cy - y);
149 if (diff < smallest) {
157 // Hone in on the specific best match around the match we found above
161 for (y = tmp2 - 1; y <= tmp2 + 1; y++) {
162 for (x = tmp - 1; x <= tmp + 1; x++) {
163 if (x == tmp && y == tmp2)
166 diff = CMP(cx - x, cy - y);
167 if (diff < smallest) {
176 if (smallest > 512) {
181 //av_log(NULL, AV_LOG_ERROR, "%d\n", smallest);
182 //av_log(NULL, AV_LOG_ERROR, "Final: (%d, %d) = %d x %d\n", cx, cy, mv->x, mv->y);
186 * Find the contrast of a given block. When searching for global motion we
187 * really only care about the high contrast blocks, so using this method we
188 * can actually skip blocks we don't care much about.
190 static int block_contrast(uint8_t *src, int x, int y, int stride, int blocksize)
196 for (i = 0; i <= blocksize * 2; i++) {
197 // We use a width of 16 here to match the sad function
198 for (j = 0; j <= 15; j++) {
199 pos = (y + i) * stride + (x + j);
200 if (src[pos] < lowest)
202 else if (src[pos] > highest) {
208 return highest - lowest;
212 * Find the rotation for a given block.
214 static double block_angle(int x, int y, int cx, int cy, IntMotionVector *shift)
218 a1 = atan2(y - cy, x - cx);
219 a2 = atan2(y - cy + shift->y, x - cx + shift->x);
223 return (diff > M_PI) ? diff - 2 * M_PI :
224 (diff < -M_PI) ? diff + 2 * M_PI :
229 * Find the estimated global motion for a scene given the most likely shift
230 * for each block in the frame. The global motion is estimated to be the
231 * same as the motion from most blocks in the frame, so if most blocks
232 * move one pixel to the right and two pixels down, this would yield a
233 * motion vector (1, -2).
235 static void find_motion(DeshakeContext *deshake, uint8_t *src1, uint8_t *src2,
236 int width, int height, int stride, Transform *t)
239 IntMotionVector mv = {0, 0};
240 int count_max_value = 0;
244 int center_x = 0, center_y = 0;
247 av_fast_malloc(&deshake->angles, &deshake->angles_size, width * height / (16 * deshake->blocksize) * sizeof(*deshake->angles));
249 // Reset counts to zero
250 for (x = 0; x < deshake->rx * 2 + 1; x++) {
251 for (y = 0; y < deshake->ry * 2 + 1; y++) {
252 deshake->counts[x][y] = 0;
257 // Find motion for every block and store the motion vector in the counts
258 for (y = deshake->ry; y < height - deshake->ry - (deshake->blocksize * 2); y += deshake->blocksize * 2) {
259 // We use a width of 16 here to match the sad function
260 for (x = deshake->rx; x < width - deshake->rx - 16; x += 16) {
261 // If the contrast is too low, just skip this block as it probably
262 // won't be very useful to us.
263 contrast = block_contrast(src2, x, y, stride, deshake->blocksize);
264 if (contrast > deshake->contrast) {
265 //av_log(NULL, AV_LOG_ERROR, "%d\n", contrast);
266 find_block_motion(deshake, src1, src2, x, y, stride, &mv);
267 if (mv.x != -1 && mv.y != -1) {
268 deshake->counts[mv.x + deshake->rx][mv.y + deshake->ry] += 1;
269 if (x > deshake->rx && y > deshake->ry)
270 deshake->angles[pos++] = block_angle(x, y, 0, 0, &mv);
282 t->angle = clean_mean(deshake->angles, pos);
283 if (t->angle < 0.001)
289 // Find the most common motion vector in the frame and use it as the gmv
290 for (y = deshake->ry * 2; y >= 0; y--) {
291 for (x = 0; x < deshake->rx * 2 + 1; x++) {
292 //av_log(NULL, AV_LOG_ERROR, "%5d ", deshake->counts[x][y]);
293 if (deshake->counts[x][y] > count_max_value) {
294 t->vec.x = x - deshake->rx;
295 t->vec.y = y - deshake->ry;
296 count_max_value = deshake->counts[x][y];
299 //av_log(NULL, AV_LOG_ERROR, "\n");
302 p_x = (center_x - width / 2.0);
303 p_y = (center_y - height / 2.0);
304 t->vec.x += (cos(t->angle)-1)*p_x - sin(t->angle)*p_y;
305 t->vec.y += sin(t->angle)*p_x + (cos(t->angle)-1)*p_y;
307 // Clamp max shift & rotation?
308 t->vec.x = av_clipf(t->vec.x, -deshake->rx * 2, deshake->rx * 2);
309 t->vec.y = av_clipf(t->vec.y, -deshake->ry * 2, deshake->ry * 2);
310 t->angle = av_clipf(t->angle, -0.1, 0.1);
312 //av_log(NULL, AV_LOG_ERROR, "%d x %d\n", avg->x, avg->y);
315 static int deshake_transform_c(AVFilterContext *ctx,
316 int width, int height, int cw, int ch,
317 const float *matrix_y, const float *matrix_uv,
318 enum InterpolateMethod interpolate,
319 enum FillMethod fill, AVFrame *in, AVFrame *out)
322 const float *matrixs[3];
323 int plane_w[3], plane_h[3];
324 matrixs[0] = matrix_y;
325 matrixs[1] = matrixs[2] = matrix_uv;
327 plane_w[1] = plane_w[2] = cw;
329 plane_h[1] = plane_h[2] = ch;
331 for (i = 0; i < 3; i++) {
332 // Transform the luma and chroma planes
333 ret = avfilter_transform(in->data[i], out->data[i], in->linesize[i], out->linesize[i],
334 plane_w[i], plane_h[i], matrixs[i], interpolate, fill);
341 static av_cold int init(AVFilterContext *ctx)
343 DeshakeContext *deshake = ctx->priv;
345 deshake->refcount = 20; // XXX: add to options?
346 deshake->blocksize /= 2;
347 deshake->blocksize = av_clip(deshake->blocksize, 4, 128);
349 if (deshake->rx % 16) {
350 av_log(ctx, AV_LOG_ERROR, "rx must be a multiple of 16\n");
351 return AVERROR_PATCHWELCOME;
354 if (deshake->filename)
355 deshake->fp = fopen(deshake->filename, "w");
357 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", 1, 104, deshake->fp);
359 // Quadword align left edge of box for MMX code, adjust width if necessary
360 // to keep right margin
361 if (deshake->cx > 0) {
362 deshake->cw += deshake->cx - (deshake->cx & ~15);
365 deshake->transform = deshake_transform_c;
367 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",
368 deshake->cx, deshake->cy, deshake->cw, deshake->ch,
369 deshake->rx, deshake->ry, deshake->edge, deshake->blocksize * 2, deshake->contrast, deshake->search);
374 static int query_formats(AVFilterContext *ctx)
376 static const enum AVPixelFormat pix_fmts[] = {
377 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV410P,
378 AV_PIX_FMT_YUV411P, AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ422P,
379 AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ440P, AV_PIX_FMT_NONE
381 AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
383 return AVERROR(ENOMEM);
384 return ff_set_common_formats(ctx, fmts_list);
387 static int config_props(AVFilterLink *link)
389 DeshakeContext *deshake = link->dst->priv;
392 deshake->last.vec.x = 0;
393 deshake->last.vec.y = 0;
394 deshake->last.angle = 0;
395 deshake->last.zoom = 0;
400 static av_cold void uninit(AVFilterContext *ctx)
402 DeshakeContext *deshake = ctx->priv;
403 av_frame_free(&deshake->ref);
404 av_freep(&deshake->angles);
405 deshake->angles_size = 0;
410 static int filter_frame(AVFilterLink *link, AVFrame *in)
412 DeshakeContext *deshake = link->dst->priv;
413 AVFilterLink *outlink = link->dst->outputs[0];
415 Transform t = {{0},0}, orig = {{0},0};
416 float matrix_y[9], matrix_uv[9];
417 float alpha = 2.0 / deshake->refcount;
420 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(link->format);
421 const int chroma_width = AV_CEIL_RSHIFT(link->w, desc->log2_chroma_w);
422 const int chroma_height = AV_CEIL_RSHIFT(link->h, desc->log2_chroma_h);
424 float transform_zoom;
426 out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
429 return AVERROR(ENOMEM);
431 av_frame_copy_props(out, in);
433 aligned = !((intptr_t)in->data[0] & 15 | in->linesize[0] & 15);
434 deshake->sad = av_pixelutils_get_sad_fn(4, 4, aligned, deshake); // 16x16, 2nd source unaligned
436 return AVERROR(EINVAL);
438 if (deshake->cx < 0 || deshake->cy < 0 || deshake->cw < 0 || deshake->ch < 0) {
439 // Find the most likely global motion for the current frame
440 find_motion(deshake, (deshake->ref == NULL) ? in->data[0] : deshake->ref->data[0], in->data[0], link->w, link->h, in->linesize[0], &t);
442 uint8_t *src1 = (deshake->ref == NULL) ? in->data[0] : deshake->ref->data[0];
443 uint8_t *src2 = in->data[0];
445 deshake->cx = FFMIN(deshake->cx, link->w);
446 deshake->cy = FFMIN(deshake->cy, link->h);
448 if ((unsigned)deshake->cx + (unsigned)deshake->cw > link->w) deshake->cw = link->w - deshake->cx;
449 if ((unsigned)deshake->cy + (unsigned)deshake->ch > link->h) deshake->ch = link->h - deshake->cy;
451 // Quadword align right margin
454 src1 += deshake->cy * in->linesize[0] + deshake->cx;
455 src2 += deshake->cy * in->linesize[0] + deshake->cx;
457 find_motion(deshake, src1, src2, deshake->cw, deshake->ch, in->linesize[0], &t);
461 // Copy transform so we can output it later to compare to the smoothed value
462 orig.vec.x = t.vec.x;
463 orig.vec.y = t.vec.y;
464 orig.angle = t.angle;
467 // Generate a one-sided moving exponential average
468 deshake->avg.vec.x = alpha * t.vec.x + (1.0 - alpha) * deshake->avg.vec.x;
469 deshake->avg.vec.y = alpha * t.vec.y + (1.0 - alpha) * deshake->avg.vec.y;
470 deshake->avg.angle = alpha * t.angle + (1.0 - alpha) * deshake->avg.angle;
471 deshake->avg.zoom = alpha * t.zoom + (1.0 - alpha) * deshake->avg.zoom;
473 // Remove the average from the current motion to detect the motion that
474 // is not on purpose, just as jitter from bumping the camera
475 t.vec.x -= deshake->avg.vec.x;
476 t.vec.y -= deshake->avg.vec.y;
477 t.angle -= deshake->avg.angle;
478 t.zoom -= deshake->avg.zoom;
480 // Invert the motion to undo it
485 // Write statistics to file
487 snprintf(tmp, 256, "%f, %f, %f, %f, %f, %f, %f, %f, %f, %f, %f, %f\n", orig.vec.x, deshake->avg.vec.x, t.vec.x, orig.vec.y, deshake->avg.vec.y, t.vec.y, orig.angle, deshake->avg.angle, t.angle, orig.zoom, deshake->avg.zoom, t.zoom);
488 fwrite(tmp, 1, strlen(tmp), deshake->fp);
491 // Turn relative current frame motion into absolute by adding it to the
492 // last absolute motion
493 t.vec.x += deshake->last.vec.x;
494 t.vec.y += deshake->last.vec.y;
495 t.angle += deshake->last.angle;
496 t.zoom += deshake->last.zoom;
498 // Shrink motion by 10% to keep things centered in the camera frame
503 // Store the last absolute motion information
504 deshake->last.vec.x = t.vec.x;
505 deshake->last.vec.y = t.vec.y;
506 deshake->last.angle = t.angle;
507 deshake->last.zoom = t.zoom;
509 transform_zoom = 1.0 + t.zoom / 100.0;
511 // Generate a luma transformation matrix
512 ff_get_matrix(t.vec.x, t.vec.y, t.angle, transform_zoom, transform_zoom, matrix_y);
513 // Generate a chroma transformation matrix
514 ff_get_matrix(t.vec.x / (link->w / chroma_width), t.vec.y / (link->h / chroma_height), t.angle, transform_zoom, transform_zoom, matrix_uv);
515 // Transform the luma and chroma planes
516 ret = deshake->transform(link->dst, link->w, link->h, chroma_width, chroma_height,
517 matrix_y, matrix_uv, INTERPOLATE_BILINEAR, deshake->edge, in, out);
519 // Cleanup the old reference frame
520 av_frame_free(&deshake->ref);
525 // Store the current frame as the reference frame for calculating the
526 // motion of the next frame
529 return ff_filter_frame(outlink, out);
535 static const AVFilterPad deshake_inputs[] = {
538 .type = AVMEDIA_TYPE_VIDEO,
539 .filter_frame = filter_frame,
540 .config_props = config_props,
545 static const AVFilterPad deshake_outputs[] = {
548 .type = AVMEDIA_TYPE_VIDEO,
553 AVFilter ff_vf_deshake = {
555 .description = NULL_IF_CONFIG_SMALL("Stabilize shaky video."),
556 .priv_size = sizeof(DeshakeContext),
559 .query_formats = query_formats,
560 .inputs = deshake_inputs,
561 .outputs = deshake_outputs,
562 .priv_class = &deshake_class,