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 "libavcodec/dsputil.h"
63 #include "deshake_opencl.h"
65 #define CHROMA_WIDTH(link) -((-link->w) >> av_pix_fmt_desc_get(link->format)->log2_chroma_w)
66 #define CHROMA_HEIGHT(link) -((-link->h) >> av_pix_fmt_desc_get(link->format)->log2_chroma_h)
68 #define OFFSET(x) offsetof(DeshakeContext, x)
69 #define FLAGS AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
71 static const AVOption deshake_options[] = {
72 { "x", "set x for the rectangular search area", OFFSET(cx), AV_OPT_TYPE_INT, {.i64=-1}, -1, INT_MAX, .flags = FLAGS },
73 { "y", "set y for the rectangular search area", OFFSET(cy), AV_OPT_TYPE_INT, {.i64=-1}, -1, INT_MAX, .flags = FLAGS },
74 { "w", "set width for the rectangular search area", OFFSET(cw), AV_OPT_TYPE_INT, {.i64=-1}, -1, INT_MAX, .flags = FLAGS },
75 { "h", "set height for the rectangular search area", OFFSET(ch), AV_OPT_TYPE_INT, {.i64=-1}, -1, INT_MAX, .flags = FLAGS },
76 { "rx", "set x for the rectangular search area", OFFSET(rx), AV_OPT_TYPE_INT, {.i64=16}, 0, 64, .flags = FLAGS },
77 { "ry", "set y for the rectangular search area", OFFSET(ry), AV_OPT_TYPE_INT, {.i64=16}, 0, 64, .flags = FLAGS },
78 { "edge", "set edge mode", OFFSET(edge), AV_OPT_TYPE_INT, {.i64=FILL_MIRROR}, FILL_BLANK, FILL_COUNT-1, FLAGS, "edge"},
79 { "blank", "fill zeroes at blank locations", 0, AV_OPT_TYPE_CONST, {.i64=FILL_BLANK}, INT_MIN, INT_MAX, FLAGS, "edge" },
80 { "original", "original image at blank locations", 0, AV_OPT_TYPE_CONST, {.i64=FILL_ORIGINAL}, INT_MIN, INT_MAX, FLAGS, "edge" },
81 { "clamp", "extruded edge value at blank locations", 0, AV_OPT_TYPE_CONST, {.i64=FILL_CLAMP}, INT_MIN, INT_MAX, FLAGS, "edge" },
82 { "mirror", "mirrored edge at blank locations", 0, AV_OPT_TYPE_CONST, {.i64=FILL_MIRROR}, INT_MIN, INT_MAX, FLAGS, "edge" },
83 { "blocksize", "set motion search blocksize", OFFSET(blocksize), AV_OPT_TYPE_INT, {.i64=8}, 4, 128, .flags = FLAGS },
84 { "contrast", "set contrast threshold for blocks", OFFSET(contrast), AV_OPT_TYPE_INT, {.i64=125}, 1, 255, .flags = FLAGS },
85 { "search", "set search strategy", OFFSET(search), AV_OPT_TYPE_INT, {.i64=EXHAUSTIVE}, EXHAUSTIVE, SEARCH_COUNT-1, FLAGS, "smode" },
86 { "exhaustive", "exhaustive search", 0, AV_OPT_TYPE_CONST, {.i64=EXHAUSTIVE}, INT_MIN, INT_MAX, FLAGS, "smode" },
87 { "less", "less exhaustive search", 0, AV_OPT_TYPE_CONST, {.i64=SMART_EXHAUSTIVE}, INT_MIN, INT_MAX, FLAGS, "smode" },
88 { "filename", "set motion search detailed log file name", OFFSET(filename), AV_OPT_TYPE_STRING, {.str=NULL}, .flags = FLAGS },
89 { "opencl", "use OpenCL filtering capabilities", OFFSET(opencl), AV_OPT_TYPE_INT, {.i64=0}, 0, 1, .flags = FLAGS },
93 AVFILTER_DEFINE_CLASS(deshake);
95 static int cmp(const double *a, const double *b)
97 return *a < *b ? -1 : ( *a > *b ? 1 : 0 );
101 * Cleaned mean (cuts off 20% of values to remove outliers and then averages)
103 static double clean_mean(double *values, int count)
109 qsort(values, count, sizeof(double), (void*)cmp);
111 for (x = cut; x < count - cut; x++) {
115 return mean / (count - cut * 2);
119 * Find the most likely shift in motion between two frames for a given
120 * macroblock. Test each block against several shifts given by the rx
121 * and ry attributes. Searches using a simple matrix of those shifts and
122 * chooses the most likely shift by the smallest difference in blocks.
124 static void find_block_motion(DeshakeContext *deshake, uint8_t *src1,
125 uint8_t *src2, int cx, int cy, int stride,
130 int smallest = INT_MAX;
133 #define CMP(i, j) deshake->c.sad[0](deshake, src1 + cy * stride + cx, \
134 src2 + (j) * stride + (i), stride, \
137 if (deshake->search == EXHAUSTIVE) {
138 // Compare every possible position - this is sloooow!
139 for (y = -deshake->ry; y <= deshake->ry; y++) {
140 for (x = -deshake->rx; x <= deshake->rx; x++) {
141 diff = CMP(cx - x, cy - y);
142 if (diff < smallest) {
149 } else if (deshake->search == SMART_EXHAUSTIVE) {
150 // Compare every other possible position and find the best match
151 for (y = -deshake->ry + 1; y < deshake->ry - 2; y += 2) {
152 for (x = -deshake->rx + 1; x < deshake->rx - 2; x += 2) {
153 diff = CMP(cx - x, cy - y);
154 if (diff < smallest) {
162 // Hone in on the specific best match around the match we found above
166 for (y = tmp2 - 1; y <= tmp2 + 1; y++) {
167 for (x = tmp - 1; x <= tmp + 1; x++) {
168 if (x == tmp && y == tmp2)
171 diff = CMP(cx - x, cy - y);
172 if (diff < smallest) {
181 if (smallest > 512) {
186 //av_log(NULL, AV_LOG_ERROR, "%d\n", smallest);
187 //av_log(NULL, AV_LOG_ERROR, "Final: (%d, %d) = %d x %d\n", cx, cy, mv->x, mv->y);
191 * Find the contrast of a given block. When searching for global motion we
192 * really only care about the high contrast blocks, so using this method we
193 * can actually skip blocks we don't care much about.
195 static int block_contrast(uint8_t *src, int x, int y, int stride, int blocksize)
201 for (i = 0; i <= blocksize * 2; i++) {
202 // We use a width of 16 here to match the libavcodec sad functions
203 for (j = 0; i <= 15; i++) {
204 pos = (y - i) * stride + (x - j);
205 if (src[pos] < lowest)
207 else if (src[pos] > highest) {
213 return highest - lowest;
217 * Find the rotation for a given block.
219 static double block_angle(int x, int y, int cx, int cy, IntMotionVector *shift)
223 a1 = atan2(y - cy, x - cx);
224 a2 = atan2(y - cy + shift->y, x - cx + shift->x);
228 return (diff > M_PI) ? diff - 2 * M_PI :
229 (diff < -M_PI) ? diff + 2 * M_PI :
234 * Find the estimated global motion for a scene given the most likely shift
235 * for each block in the frame. The global motion is estimated to be the
236 * same as the motion from most blocks in the frame, so if most blocks
237 * move one pixel to the right and two pixels down, this would yield a
238 * motion vector (1, -2).
240 static void find_motion(DeshakeContext *deshake, uint8_t *src1, uint8_t *src2,
241 int width, int height, int stride, Transform *t)
244 IntMotionVector mv = {0, 0};
245 int counts[128][128];
246 int count_max_value = 0;
250 double *angles = av_malloc(sizeof(*angles) * width * height / (16 * deshake->blocksize));
251 int center_x = 0, center_y = 0;
254 // Reset counts to zero
255 for (x = 0; x < deshake->rx * 2 + 1; x++) {
256 for (y = 0; y < deshake->ry * 2 + 1; y++) {
262 // Find motion for every block and store the motion vector in the counts
263 for (y = deshake->ry; y < height - deshake->ry - (deshake->blocksize * 2); y += deshake->blocksize * 2) {
264 // We use a width of 16 here to match the libavcodec sad functions
265 for (x = deshake->rx; x < width - deshake->rx - 16; x += 16) {
266 // If the contrast is too low, just skip this block as it probably
267 // won't be very useful to us.
268 contrast = block_contrast(src2, x, y, stride, deshake->blocksize);
269 if (contrast > deshake->contrast) {
270 //av_log(NULL, AV_LOG_ERROR, "%d\n", contrast);
271 find_block_motion(deshake, src1, src2, x, y, stride, &mv);
272 if (mv.x != -1 && mv.y != -1) {
273 counts[mv.x + deshake->rx][mv.y + deshake->ry] += 1;
274 if (x > deshake->rx && y > deshake->ry)
275 angles[pos++] = block_angle(x, y, 0, 0, &mv);
287 t->angle = clean_mean(angles, pos);
288 if (t->angle < 0.001)
294 // Find the most common motion vector in the frame and use it as the gmv
295 for (y = deshake->ry * 2; y >= 0; y--) {
296 for (x = 0; x < deshake->rx * 2 + 1; x++) {
297 //av_log(NULL, AV_LOG_ERROR, "%5d ", counts[x][y]);
298 if (counts[x][y] > count_max_value) {
299 t->vector.x = x - deshake->rx;
300 t->vector.y = y - deshake->ry;
301 count_max_value = counts[x][y];
304 //av_log(NULL, AV_LOG_ERROR, "\n");
307 p_x = (center_x - width / 2);
308 p_y = (center_y - height / 2);
309 t->vector.x += (cos(t->angle)-1)*p_x - sin(t->angle)*p_y;
310 t->vector.y += sin(t->angle)*p_x + (cos(t->angle)-1)*p_y;
312 // Clamp max shift & rotation?
313 t->vector.x = av_clipf(t->vector.x, -deshake->rx * 2, deshake->rx * 2);
314 t->vector.y = av_clipf(t->vector.y, -deshake->ry * 2, deshake->ry * 2);
315 t->angle = av_clipf(t->angle, -0.1, 0.1);
317 //av_log(NULL, AV_LOG_ERROR, "%d x %d\n", avg->x, avg->y);
321 static int deshake_transform_c(AVFilterContext *ctx,
322 int width, int height, int cw, int ch,
323 const float *matrix_y, const float *matrix_uv,
324 enum InterpolateMethod interpolate,
325 enum FillMethod fill, AVFrame *in, AVFrame *out)
328 const float *matrixs[3];
329 int plane_w[3], plane_h[3];
330 matrixs[0] = matrix_y;
331 matrixs[1] = matrixs[2] = matrix_uv;
333 plane_w[1] = plane_w[2] = cw;
335 plane_h[1] = plane_h[2] = ch;
337 for (i = 0; i < 3; i++) {
338 // Transform the luma and chroma planes
339 ret = avfilter_transform(in->data[i], out->data[i], in->linesize[i], out->linesize[i],
340 plane_w[i], plane_h[i], matrixs[i], interpolate, fill);
347 static av_cold int init(AVFilterContext *ctx, const char *args)
350 DeshakeContext *deshake = ctx->priv;
352 deshake->refcount = 20; // XXX: add to options?
353 deshake->blocksize /= 2;
354 deshake->blocksize = av_clip(deshake->blocksize, 4, 128);
356 if (deshake->filename)
357 deshake->fp = fopen(deshake->filename, "w");
359 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);
361 // Quadword align left edge of box for MMX code, adjust width if necessary
362 // to keep right margin
363 if (deshake->cx > 0) {
364 deshake->cw += deshake->cx - (deshake->cx & ~15);
367 deshake->transform = deshake_transform_c;
368 if (!CONFIG_OPENCL && deshake->opencl) {
369 av_log(ctx, AV_LOG_ERROR, "OpenCL support was not enabled in this build, cannot be selected\n");
370 return AVERROR(EINVAL);
373 if (deshake->opencl && CONFIG_OPENCL) {
374 deshake->transform = ff_opencl_transform;
375 ret = ff_opencl_deshake_init(ctx);
379 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",
380 deshake->cx, deshake->cy, deshake->cw, deshake->ch,
381 deshake->rx, deshake->ry, deshake->edge, deshake->blocksize * 2, deshake->contrast, deshake->search);
386 static int query_formats(AVFilterContext *ctx)
388 static const enum AVPixelFormat pix_fmts[] = {
389 AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV410P,
390 AV_PIX_FMT_YUV411P, AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ422P,
391 AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ440P, AV_PIX_FMT_NONE
394 ff_set_common_formats(ctx, ff_make_format_list(pix_fmts));
399 static int config_props(AVFilterLink *link)
401 DeshakeContext *deshake = link->dst->priv;
404 deshake->last.vector.x = 0;
405 deshake->last.vector.y = 0;
406 deshake->last.angle = 0;
407 deshake->last.zoom = 0;
409 deshake->avctx = avcodec_alloc_context3(NULL);
410 dsputil_init(&deshake->c, deshake->avctx);
415 static av_cold void uninit(AVFilterContext *ctx)
417 DeshakeContext *deshake = ctx->priv;
418 if (deshake->opencl && CONFIG_OPENCL) {
419 ff_opencl_deshake_uninit(ctx);
421 av_frame_free(&deshake->ref);
425 avcodec_close(deshake->avctx);
426 av_freep(&deshake->avctx);
429 static int filter_frame(AVFilterLink *link, AVFrame *in)
431 DeshakeContext *deshake = link->dst->priv;
432 AVFilterLink *outlink = link->dst->outputs[0];
434 Transform t = {{0},0}, orig = {{0},0};
435 float matrix_y[9], matrix_uv[9];
436 float alpha = 2.0 / deshake->refcount;
440 out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
443 return AVERROR(ENOMEM);
445 av_frame_copy_props(out, in);
447 if (deshake->opencl && CONFIG_OPENCL) {
448 ret = ff_opencl_deshake_process_inout_buf(link->dst,in, out);
453 if (deshake->cx < 0 || deshake->cy < 0 || deshake->cw < 0 || deshake->ch < 0) {
454 // Find the most likely global motion for the current frame
455 find_motion(deshake, (deshake->ref == NULL) ? in->data[0] : deshake->ref->data[0], in->data[0], link->w, link->h, in->linesize[0], &t);
457 uint8_t *src1 = (deshake->ref == NULL) ? in->data[0] : deshake->ref->data[0];
458 uint8_t *src2 = in->data[0];
460 deshake->cx = FFMIN(deshake->cx, link->w);
461 deshake->cy = FFMIN(deshake->cy, link->h);
463 if ((unsigned)deshake->cx + (unsigned)deshake->cw > link->w) deshake->cw = link->w - deshake->cx;
464 if ((unsigned)deshake->cy + (unsigned)deshake->ch > link->h) deshake->ch = link->h - deshake->cy;
466 // Quadword align right margin
469 src1 += deshake->cy * in->linesize[0] + deshake->cx;
470 src2 += deshake->cy * in->linesize[0] + deshake->cx;
472 find_motion(deshake, src1, src2, deshake->cw, deshake->ch, in->linesize[0], &t);
476 // Copy transform so we can output it later to compare to the smoothed value
477 orig.vector.x = t.vector.x;
478 orig.vector.y = t.vector.y;
479 orig.angle = t.angle;
482 // Generate a one-sided moving exponential average
483 deshake->avg.vector.x = alpha * t.vector.x + (1.0 - alpha) * deshake->avg.vector.x;
484 deshake->avg.vector.y = alpha * t.vector.y + (1.0 - alpha) * deshake->avg.vector.y;
485 deshake->avg.angle = alpha * t.angle + (1.0 - alpha) * deshake->avg.angle;
486 deshake->avg.zoom = alpha * t.zoom + (1.0 - alpha) * deshake->avg.zoom;
488 // Remove the average from the current motion to detect the motion that
489 // is not on purpose, just as jitter from bumping the camera
490 t.vector.x -= deshake->avg.vector.x;
491 t.vector.y -= deshake->avg.vector.y;
492 t.angle -= deshake->avg.angle;
493 t.zoom -= deshake->avg.zoom;
495 // Invert the motion to undo it
500 // Write statistics to file
502 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);
503 fwrite(tmp, sizeof(char), strlen(tmp), deshake->fp);
506 // Turn relative current frame motion into absolute by adding it to the
507 // last absolute motion
508 t.vector.x += deshake->last.vector.x;
509 t.vector.y += deshake->last.vector.y;
510 t.angle += deshake->last.angle;
511 t.zoom += deshake->last.zoom;
513 // Shrink motion by 10% to keep things centered in the camera frame
518 // Store the last absolute motion information
519 deshake->last.vector.x = t.vector.x;
520 deshake->last.vector.y = t.vector.y;
521 deshake->last.angle = t.angle;
522 deshake->last.zoom = t.zoom;
524 // Generate a luma transformation matrix
525 avfilter_get_matrix(t.vector.x, t.vector.y, t.angle, 1.0 + t.zoom / 100.0, matrix_y);
526 // Generate a chroma transformation matrix
527 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_uv);
528 // Transform the luma and chroma planes
529 ret = deshake->transform(link->dst, link->w, link->h, CHROMA_WIDTH(link), CHROMA_HEIGHT(link),
530 matrix_y, matrix_uv, INTERPOLATE_BILINEAR, deshake->edge, in, out);
532 // Cleanup the old reference frame
533 av_frame_free(&deshake->ref);
538 // Store the current frame as the reference frame for calculating the
539 // motion of the next frame
542 return ff_filter_frame(outlink, out);
545 static const AVFilterPad deshake_inputs[] = {
548 .type = AVMEDIA_TYPE_VIDEO,
549 .filter_frame = filter_frame,
550 .config_props = config_props,
555 static const AVFilterPad deshake_outputs[] = {
558 .type = AVMEDIA_TYPE_VIDEO,
563 static const char *const shorthand[] = {
564 "x", "y", "w", "h", "rx", "ry", "edge",
565 "blocksize", "contrast", "search", "filename",
569 AVFilter avfilter_vf_deshake = {
571 .description = NULL_IF_CONFIG_SMALL("Stabilize shaky video."),
572 .priv_size = sizeof(DeshakeContext),
575 .query_formats = query_formats,
576 .inputs = deshake_inputs,
577 .outputs = deshake_outputs,
578 .priv_class = &deshake_class,
579 .shorthand = shorthand,