2 * /brief fast motion estimation filter
3 * /author Zachary Drew, Copyright 2005
5 * Currently only uses Gamma data for comparisonon (bug or feature?)
6 * SSE optimized where available.
8 * Vector orientation: The vector data that is generated for the current frame specifies
9 * the motion from the previous frame to the current frame. To know how a macroblock
10 * in the current frame will move in the future, the next frame is needed.
12 * This program is free software; you can redistribute it and/or modify
13 * it under the terms of the GNU General Public License as published by
14 * the Free Software Foundation; either version 2 of the License, or
15 * (at your option) any later version.
17 * This program is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
20 * GNU General Public License for more details.
22 * You should have received a copy of the GNU General Public License
23 * along with this program; if not, write to the Free Software Foundation,
24 * Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
28 #include "filter_motion_est.h"
29 #include <framework/mlt.h>
49 #define DIAMOND_SEARCH 0x0
50 #define FULL_SEARCH 0x1
52 #define MIN(a,b) ((a) > (b) ? (b) : (a))
53 #define ABS(a) ((a) >= 0 ? (a) : (-(a)))
56 struct motion_est_context_s
58 int initialized; // true if filter has been initialized
64 /* same as mlt_frame's parameters */
67 /* Operational details */
70 uint8_t *cache_image; // Copy of current frame
71 uint8_t *former_image; // Copy of former frame
75 int limit_x, limit_y; // max x and y of a motion vector
77 int check_chroma; // if check_chroma == 1 then compare chroma
80 int show_reconstruction;
81 int toggle_when_paused;
85 struct mlt_geometry_item_s bounds; // Current bounds (from filters crop_detect, autotrack rectangle, or other)
87 /* bounds in macroblock units; macroblocks are completely contained within the boundry */
88 int left_mb, prev_left_mb, right_mb, prev_right_mb;
89 int top_mb, prev_top_mb, bottom_mb, prev_bottom_mb;
91 /* size of our vector buffers */
92 int mv_buffer_height, mv_buffer_width, mv_size;
95 int former_vectors_valid; //<! true if the previous frame's buffered motion vector data is valid
96 motion_vector *former_vectors;
97 motion_vector *current_vectors;
98 motion_vector *denoise_vectors;
99 mlt_position former_frame_position, current_frame_position;
101 /* diagnostic metrics */
102 float predictive_misses; // How often do the prediction motion vectors fail?
103 int comparison_average; // How far does the best estimation deviate from a perfect comparison?
106 int average_x, average_y;
108 /* run-time configurable comparison functions */
109 int (*compare_reference)(uint8_t *, uint8_t *, int, int, int, int);
110 int (*compare_optimized)(uint8_t *, uint8_t *, int, int, int, int);
114 // This is used to constrains pixel operations between two blocks to be within the image boundry
115 inline static int constrain( int *x, int *y, int *w, int *h,
116 const int dx, const int dy,
117 const int left, const int right,
118 const int top, const int bottom)
120 uint32_t penalty = 1 << SHIFT; // Retain a few extra bits of precision
126 // Origin of macroblock moves left of image boundy
127 if( *x < left || x2 < left ) {
128 w_remains = *w - left + ((*x < x2) ? *x : x2);
129 *x += *w - w_remains;
131 // Portion of macroblock moves right of image boundry
132 else if( *x + *w > right || x2 + *w > right )
133 w_remains = right - ((*x > x2) ? *x : x2);
135 // Origin of macroblock moves above image boundy
136 if( *y < top || y2 < top ) {
137 h_remains = *h - top + ((*y < y2) ? *y : y2);
138 *y += *h - h_remains;
140 // Portion of macroblock moves bellow image boundry
141 else if( *y + *h > bottom || y2 + *h > bottom )
142 h_remains = bottom - ((*y > y2) ? *y : y2);
144 if( w_remains == *w && h_remains == *h ) return penalty;
145 if( w_remains <= 0 || h_remains <= 0) return 0; // Block is clipped out of existance
146 penalty = (*w * *h * penalty)
147 / ( w_remains * h_remains); // Recipricol of the fraction of the block that remains
149 assert(*x >= left); assert(x2 + *w - w_remains >= left);
150 assert(*y >= top); assert(y2 + *h - h_remains >= top);
151 assert(*x + w_remains <= right); assert(x2 + w_remains <= right);
152 assert(*y + h_remains <= bottom); assert(y2 + h_remains <= bottom);
154 *w = w_remains; // Update the width and height
160 /** /brief Reference Sum of Absolute Differences comparison function
163 static int sad_reference( uint8_t *block1, uint8_t *block2, const int xstride, const int ystride, const int w, const int h )
166 for ( j = 0; j < h; j++ ){
167 for ( i = 0; i < w; i++ ){
168 score += ABS( block1[i*xstride] - block2[i*xstride] );
178 /** /brief Abstracted block comparison function
180 inline static int block_compare( uint8_t *block1,
186 struct motion_est_context_s *c)
189 #ifdef COUNT_COMPARES
195 // Default comparison may be overridden by the slower, more capable reference comparison
196 int (*cmp)(uint8_t *, uint8_t *, int, int, int, int) = c->compare_optimized;
198 // vector displacement limited has been exceeded
199 if( ABS( dx ) >= c->limit_x || ABS( dy ) >= c->limit_y )
202 int mb_w = c->mb_w; // Some writeable local copies
205 // Determine if either macroblock got clipped
206 int penalty = constrain( &x, &y, &mb_w, &mb_h, dx, dy, 0, c->width, 0, c->height);
209 if( penalty == 0 ) // Clipped out of existance: Return worst score
211 else if( penalty != 1<<SHIFT ) // Nonstandard macroblock dimensions: Disable SIMD optimizizations.
212 cmp = c->compare_reference;
214 // Calculate the memory locations of the macroblocks
215 block1 += x * c->xstride + y * c->ystride;
216 block2 += (x+dx) * c->xstride + (y+dy) * c->ystride;
219 if( penalty == 1<<SHIFT ){
220 score = c->compare_reference( block1, block2, c->xstride, c->ystride, mb_w, mb_h );
221 int score2 = c->compare_optimized( block1, block2, c->xstride, c->ystride, mb_w, mb_h );
222 if ( score != score2 )
223 fprintf(stderr, "Your assembly doesn't work! Reference: %d Asm: %d\n", score, score2);
228 score = cmp( block1, block2, c->xstride, c->ystride, mb_w, mb_h );
230 return ( score * penalty ) >> SHIFT; // Ditch the extra precision
233 static inline void check_candidates ( uint8_t *ref,
234 uint8_t *candidate_base,
237 const motion_vector *candidates,// Contains to_x & to_y
238 const int count, // Number of candidates
239 const int unique, // Sometimes we know the candidates are unique
240 motion_vector *result,
241 struct motion_est_context_s *c )
244 /* Scan for the best candidate */
245 for ( i = 0; i < count; i++ )
247 // this little dohicky ignores duplicate candidates, if they are possible
252 if ( candidates[j].dx == candidates[i].dx &&
253 candidates[j].dy == candidates[i].dy )
261 score = block_compare( ref, candidate_base,
263 candidates[i].dx, // from
267 if ( score < result->msad ) { // New minimum
268 result->dx = candidates[i].dx;
269 result->dy = candidates[i].dy;
270 result->msad = score;
276 /* /brief Diamond search
277 * Operates on a single macroblock
279 static inline void diamond_search(
280 uint8_t *ref, //<! Image data from previous frame
281 uint8_t *candidate_base, //<! Image data in current frame
282 const int x, //<! X upper left corner of macroblock
283 const int y, //<! U upper left corner of macroblock
284 struct motion_vector_s *result, //<! Best predicted mv and eventual result
285 struct motion_est_context_s *c) //<! motion estimation context
288 // diamond search pattern
289 motion_vector candidates[4];
291 // Keep track of best and former best candidates
292 motion_vector best, former;
298 // The direction of the refinement needs to be known
299 motion_vector current;
303 // Loop through the search pattern
306 current.dx = result->dx;
307 current.dy = result->dy;
309 if ( first == 1 ) // Set the initial pattern
311 candidates[0].dx = result->dx + 1; candidates[0].dy = result->dy + 0;
312 candidates[1].dx = result->dx + 0; candidates[1].dy = result->dy + 1;
313 candidates[2].dx = result->dx - 1; candidates[2].dy = result->dy + 0;
314 candidates[3].dx = result->dx + 0; candidates[3].dy = result->dy - 1;
317 else // Construct the next portion of the search pattern
319 candidates[0].dx = result->dx + best.dx;
320 candidates[0].dy = result->dy + best.dy;
321 if (best.dx == former.dx && best.dy == former.dy) {
322 candidates[1].dx = result->dx + best.dy;
323 candidates[1].dy = result->dy + best.dx; // Yes, the wires
324 candidates[2].dx = result->dx - best.dy; // are crossed
325 candidates[2].dy = result->dy - best.dx;
328 candidates[1].dx = result->dx + former.dx;
329 candidates[1].dy = result->dy + former.dy;
333 former.dx = best.dx; former.dy = best.dy; // Keep track of new former best
336 check_candidates ( ref, candidate_base, x, y, candidates, i, 1, result, c );
338 // Which candidate was the best?
339 best.dx = result->dx - current.dx;
340 best.dy = result->dy - current.dy;
342 // A better canidate was not found
343 if ( best.dx == 0 && best.dy == 0 )
348 former.dx = best.dx; former.dy = best.dy; // First iteration, sensible value for former.d*
353 /* /brief Full (brute) search
354 * Operates on a single macroblock
356 __attribute__((used))
357 static void full_search(
358 uint8_t *ref, //<! Image data from previous frame
359 uint8_t *candidate_base, //<! Image data in current frame
360 int x, //<! X upper left corner of macroblock
361 int y, //<! U upper left corner of macroblock
362 struct motion_vector_s *result, //<! Best predicted mv and eventual result
363 struct motion_est_context_s *c) //<! motion estimation context
365 // Keep track of best candidate
369 for( i = -c->mb_w; i <= c->mb_w; i++ ){
370 for( j = -c->mb_h; j <= c->mb_h; j++ ){
372 score = block_compare( ref, candidate_base,
379 if ( score < result->msad ) {
382 result->msad = score;
388 // Macros for pointer calculations
389 #define CURRENT(i,j) ( c->current_vectors + (j)*c->mv_buffer_width + (i) )
390 #define FORMER(i,j) ( c->former_vectors + (j)*c->mv_buffer_width + (i) )
391 #define DENOISE(i,j) ( c->denoise_vectors + (j)*c->mv_buffer_width + (i) )
393 int ncompare (const void * a, const void * b)
395 return ( *(const int*)a - *(const int*)b );
398 // motion vector denoising
399 // for x and y components seperately,
400 // change the vector to be the median value of the 9 adjacent vectors
401 static void median_denoise( motion_vector *v, struct motion_est_context_s *c )
403 int xvalues[9], yvalues[9];
406 for( j = c->top_mb; j <= c->bottom_mb; j++ )
407 for( i = c->left_mb; i <= c->right_mb; i++ ){
411 xvalues[n ] = CURRENT(i,j)->dx; // Center
412 yvalues[n++] = CURRENT(i,j)->dy;
414 if( i > c->left_mb ) // Not in First Column
416 xvalues[n ] = CURRENT(i-1,j)->dx; // Left
417 yvalues[n++] = CURRENT(i-1,j)->dy;
419 if( j > c->top_mb ) {
420 xvalues[n ] = CURRENT(i-1,j-1)->dx; // Upper Left
421 yvalues[n++] = CURRENT(i-1,j-1)->dy;
424 if( j < c->bottom_mb ) {
425 xvalues[n ] = CURRENT(i-1,j+1)->dx; // Bottom Left
426 yvalues[n++] = CURRENT(i-1,j+1)->dy;
429 if( i < c->right_mb ) // Not in Last Column
431 xvalues[n ] = CURRENT(i+1,j)->dx; // Right
432 yvalues[n++] = CURRENT(i+1,j)->dy;
435 if( j > c->top_mb ) {
436 xvalues[n ] = CURRENT(i+1,j-1)->dx; // Upper Right
437 yvalues[n++] = CURRENT(i+1,j-1)->dy;
440 if( j < c->bottom_mb ) {
441 xvalues[n ] = CURRENT(i+1,j+1)->dx; // Bottom Right
442 yvalues[n++] = CURRENT(i+1,j+1)->dy;
445 if( j > c->top_mb ) // Not in First Row
447 xvalues[n ] = CURRENT(i,j-1)->dx; // Top
448 yvalues[n++] = CURRENT(i,j-1)->dy;
451 if( j < c->bottom_mb ) // Not in Last Row
453 xvalues[n ] = CURRENT(i,j+1)->dx; // Bottom
454 yvalues[n++] = CURRENT(i,j+1)->dy;
457 qsort (xvalues, n, sizeof(int), ncompare);
458 qsort (yvalues, n, sizeof(int), ncompare);
461 DENOISE(i,j)->dx = xvalues[n/2];
462 DENOISE(i,j)->dy = yvalues[n/2];
465 DENOISE(i,j)->dx = (xvalues[n/2] + xvalues[n/2+1])/2;
466 DENOISE(i,j)->dy = (yvalues[n/2] + yvalues[n/2+1])/2;
471 motion_vector *t = c->current_vectors;
472 c->current_vectors = c->denoise_vectors;
473 c->denoise_vectors = t;
479 static inline int median_predictor(int a, int b, int c) {
495 /** /brief Motion search
497 * For each macroblock in the current frame, estimate the block from the last frame that
500 * Vocab: Colocated - the pixel in the previous frame at the current position
502 * Based on enhanced predictive zonal search. [Tourapis 2002]
504 static void motion_search( uint8_t *from, //<! Image data.
505 uint8_t *to, //<! Image data. Rigid grid.
506 struct motion_est_context_s *c) //<! The context
509 #ifdef COUNT_COMPARES
513 motion_vector candidates[10];
514 motion_vector *here; // This one gets used alot (about 30 times per macroblock)
519 // For every macroblock, perform motion vector estimation
520 for( i = c->left_mb; i <= c->right_mb; i++ ){
521 for( j = c->top_mb; j <= c->bottom_mb; j++ ){
526 here->msad = MAX_MSAD;
531 /* Stack the predictors [i.e. checked in reverse order] */
533 /* Adjacent to collocated */
534 if( c->former_vectors_valid )
537 if( j > c->prev_top_mb ){// && COL_TOP->valid ){
538 candidates[n ].dx = FORMER(i,j-1)->dx;
539 candidates[n++].dy = FORMER(i,j-1)->dy;
543 if( i > c->prev_left_mb ){// && COL_LEFT->valid ){
544 candidates[n ].dx = FORMER(i-1,j)->dx;
545 candidates[n++].dy = FORMER(i-1,j)->dy;
548 // Right of colocated
549 if( i < c->prev_right_mb ){// && COL_RIGHT->valid ){
550 candidates[n ].dx = FORMER(i+1,j)->dx;
551 candidates[n++].dy = FORMER(i+1,j)->dy;
554 // Bottom of colocated
555 if( j < c->prev_bottom_mb ){// && COL_BOTTOM->valid ){
556 candidates[n ].dx = FORMER(i,j+1)->dx;
557 candidates[n++].dy = FORMER(i,j+1)->dy;
560 // And finally, colocated
561 candidates[n ].dx = FORMER(i,j)->dx;
562 candidates[n++].dy = FORMER(i,j)->dy;
565 // For macroblocks not in the top row
566 if ( j > c->top_mb) {
568 // Top if ( TOP->valid ) {
569 candidates[n ].dx = CURRENT(i,j-1)->dx;
570 candidates[n++].dy = CURRENT(i,j-1)->dy;
573 // Top-Right, macroblocks not in the right row
574 if ( i < c->right_mb ){// && TOP_RIGHT->valid ) {
575 candidates[n ].dx = CURRENT(i+1,j-1)->dx;
576 candidates[n++].dy = CURRENT(i+1,j-1)->dy;
580 // Left, Macroblocks not in the left column
581 if ( i > c->left_mb ){// && LEFT->valid ) {
582 candidates[n ].dx = CURRENT(i-1,j)->dx;
583 candidates[n++].dy = CURRENT(i-1,j)->dy;
586 /* Median predictor vector (median of left, top, and top right adjacent vectors) */
587 if ( i > c->left_mb && j > c->top_mb && i < c->right_mb
588 )//&& LEFT->valid && TOP->valid && TOP_RIGHT->valid )
590 candidates[n ].dx = median_predictor( CURRENT(i-1,j)->dx, CURRENT(i,j-1)->dx, CURRENT(i+1,j-1)->dx);
591 candidates[n++].dy = median_predictor( CURRENT(i-1,j)->dy, CURRENT(i,j-1)->dy, CURRENT(i+1,j-1)->dy);
595 candidates[n ].dx = 0;
596 candidates[n++].dy = 0;
600 check_candidates ( to, from, x, y, candidates, n, 0, here, c );
604 diamond_search( to, from, x, y, here, c);
606 full_search( to, from, x, y, here, c);
609 assert( x + c->mb_w + here->dx > 0 ); // All macroblocks must have area > 0
610 assert( y + c->mb_h + here->dy > 0 );
611 assert( x + here->dx < c->width );
612 assert( y + here->dy < c->height );
614 } /* End column loop */
618 asm volatile ( "emms" );
621 #ifdef COUNT_COMPARES
622 fprintf(stderr, "%d comparisons per block were made", compares/count);
627 void collect_post_statistics( struct motion_est_context_s *c ) {
629 c->comparison_average = 0;
630 c->average_length = 0;
636 for ( i = c->left_mb; i <= c->right_mb; i++ ){
637 for ( j = c->top_mb; j <= c->bottom_mb; j++ ){
640 c->comparison_average += CURRENT(i,j)->msad;
641 c->average_x += CURRENT(i,j)->dx;
642 c->average_y += CURRENT(i,j)->dy;
650 c->comparison_average /= count;
651 c->average_x /= count;
652 c->average_y /= count;
653 c->average_length = sqrt( c->average_x * c->average_x + c->average_y * c->average_y );
658 static void init_optimizations( struct motion_est_context_s *c )
662 case 4: if(c->mb_h == 4) c->compare_optimized = sad_sse_422_luma_4x4;
663 else c->compare_optimized = sad_sse_422_luma_4w;
665 case 8: if(c->mb_h == 8) c->compare_optimized = sad_sse_422_luma_8x8;
666 else c->compare_optimized = sad_sse_422_luma_8w;
668 case 16: if(c->mb_h == 16) c->compare_optimized = sad_sse_422_luma_16x16;
669 else c->compare_optimized = sad_sse_422_luma_16w;
671 case 32: if(c->mb_h == 32) c->compare_optimized = sad_sse_422_luma_32x32;
672 else c->compare_optimized = sad_sse_422_luma_32w;
674 case 64: c->compare_optimized = sad_sse_422_luma_64w;
677 default: c->compare_optimized = sad_reference;
682 inline static void set_red(uint8_t *image, struct motion_est_context_s *c)
685 for( n = 0; n < c->width * c->height * 2; n+=4 )
695 static void show_residual( uint8_t *result, struct motion_est_context_s *c )
703 // set_red(result,c);
705 for( j = c->top_mb; j <= c->bottom_mb; j++ ){
706 for( i = c->left_mb; i <= c->right_mb; i++ ){
708 dx = CURRENT(i,j)->dx;
709 dy = CURRENT(i,j)->dy;
715 // Denoise function caused some blocks to be completely clipped, ignore them
716 if (constrain( &x, &y, &w, &h, dx, dy, 0, c->width, 0, c->height) == 0 )
719 for( ty = y; ty < y + h ; ty++ ){
720 for( tx = x; tx < x + w ; tx++ ){
722 b = c->former_image + (tx+dx)*c->xstride + (ty+dy)*c->ystride;
723 r = result + tx*c->xstride + ty*c->ystride;
725 r[0] = 16 + ABS( r[0] - b[0] );
728 r[1] = 128 + ABS( r[1] - b[1] );
730 // FIXME: may exceed boundies
731 r[1] = 128 + ABS( r[1] - ( *(b-1) + b[3] ) /2 );
738 static void show_reconstruction( uint8_t *result, struct motion_est_context_s *c )
746 for( i = c->left_mb; i <= c->right_mb; i++ ){
747 for( j = c->top_mb; j <= c->bottom_mb; j++ ){
749 dx = CURRENT(i,j)->dx;
750 dy = CURRENT(i,j)->dy;
756 // Denoise function caused some blocks to be completely clipped, ignore them
757 if (constrain( &x, &y, &w, &h, dx, dy, 0, c->width, 0, c->height) == 0 )
760 for( ty = y; ty < y + h ; ty++ ){
761 for( tx = x; tx < x + w ; tx++ ){
763 r = result + tx*c->xstride + ty*c->ystride;
764 s = c->former_image + (tx+dx)*c->xstride + (ty+dy)*c->ystride;
771 // FIXME: may exceed boundies
772 r[1] = ( *(s-1) + s[3] ) /2;
779 // Image stack(able) method
780 static int filter_get_image( mlt_frame frame, uint8_t **image, mlt_image_format *format, int *width, int *height, int writable )
783 mlt_filter filter = mlt_frame_pop_service( frame );
785 // Get the motion_est context object
786 struct motion_est_context_s *c = mlt_properties_get_data( MLT_FILTER_PROPERTIES( filter ), "context", NULL);
789 // Get the new image and frame number
790 *format = mlt_image_yuv422;
791 int error = mlt_frame_get_image( frame, image, format, width, height, 1 );
794 struct timeval start; gettimeofday(&start, NULL );
799 mlt_properties_debug( MLT_FRAME_PROPERTIES(frame), "error after mlt_frame_get_image() in motion_est", stderr );
801 c->current_frame_position = mlt_frame_get_position( frame );
803 /* Context Initialization */
804 if ( c->initialized == 0 ) {
806 // Get the filter properties object
807 mlt_properties properties = mlt_filter_properties( filter );
812 /* Get parameters that may have been overridden */
813 if( mlt_properties_get( properties, "macroblock_width") != NULL )
814 c->mb_w = mlt_properties_get_int( properties, "macroblock_width");
816 if( mlt_properties_get( properties, "macroblock_height") != NULL )
817 c->mb_h = mlt_properties_get_int( properties, "macroblock_height");
819 if( mlt_properties_get( properties, "prediction_thresh") != NULL )
820 c->initial_thresh = mlt_properties_get_int( properties, "prediction_thresh" );
822 c->initial_thresh = c->mb_w * c->mb_h;
824 if( mlt_properties_get( properties, "search_method") != NULL )
825 c->search_method = mlt_properties_get_int( properties, "search_method");
827 if( mlt_properties_get( properties, "skip_prediction") != NULL )
828 c->skip_prediction = mlt_properties_get_int( properties, "skip_prediction");
830 if( mlt_properties_get( properties, "limit_x") != NULL )
831 c->limit_x = mlt_properties_get_int( properties, "limit_x");
833 if( mlt_properties_get( properties, "limit_y") != NULL )
834 c->limit_y = mlt_properties_get_int( properties, "limit_y");
836 if( mlt_properties_get( properties, "check_chroma" ) != NULL )
837 c->check_chroma = mlt_properties_get_int( properties, "check_chroma" );
839 if( mlt_properties_get( properties, "denoise" ) != NULL )
840 c->denoise = mlt_properties_get_int( properties, "denoise" );
842 if( mlt_properties_get( properties, "show_reconstruction" ) != NULL )
843 c->show_reconstruction = mlt_properties_get_int( properties, "show_reconstruction" );
845 if( mlt_properties_get( properties, "show_residual" ) != NULL )
846 c->show_residual = mlt_properties_get_int( properties, "show_residual" );
848 if( mlt_properties_get( properties, "toggle_when_paused" ) != NULL )
849 c->toggle_when_paused = mlt_properties_get_int( properties, "toggle_when_paused" );
851 init_optimizations( c );
853 // Calculate the dimensions in macroblock units
854 c->mv_buffer_width = (*width / c->mb_w);
855 c->mv_buffer_height = (*height / c->mb_h);
857 // Size of the motion vector buffer
858 c->mv_size = c->mv_buffer_width * c->mv_buffer_height * sizeof(struct motion_vector_s);
860 // Allocate the motion vector buffers
861 c->former_vectors = mlt_pool_alloc( c->mv_size );
862 c->current_vectors = mlt_pool_alloc( c->mv_size );
863 c->denoise_vectors = mlt_pool_alloc( c->mv_size );
865 // Register motion buffers for destruction
866 mlt_properties_set_data( properties, "current_motion_vectors", (void *)c->current_vectors, 0, mlt_pool_release, NULL );
867 mlt_properties_set_data( properties, "former_motion_vectors", (void *)c->former_vectors, 0, mlt_pool_release, NULL );
868 mlt_properties_set_data( properties, "denoise_motion_vectors", (void *)c->denoise_vectors, 0, mlt_pool_release, NULL );
870 c->former_vectors_valid = 0;
871 memset( c->former_vectors, 0, c->mv_size );
874 c->ystride = c->xstride * *width;
876 // Allocate a cache for the previous frame's image
877 c->former_image = mlt_pool_alloc( *width * *height * 2 );
878 c->cache_image = mlt_pool_alloc( *width * *height * 2 );
880 // Register for destruction
881 mlt_properties_set_data( properties, "cache_image", (void *)c->cache_image, 0, mlt_pool_release, NULL );
882 mlt_properties_set_data( properties, "former_image", (void *)c->former_image, 0, mlt_pool_release, NULL );
884 c->former_frame_position = c->current_frame_position;
885 c->previous_msad = 0;
890 /* Check to see if somebody else has given us bounds */
891 struct mlt_geometry_item_s *bounds = mlt_properties_get_data( MLT_FRAME_PROPERTIES( frame ), "bounds", NULL );
893 if( bounds != NULL ) {
894 // translate pixel units (from bounds) to macroblock units
895 // make sure whole macroblock stays within bounds
896 c->left_mb = ( bounds->x + c->mb_w - 1 ) / c->mb_w;
897 c->top_mb = ( bounds->y + c->mb_h - 1 ) / c->mb_h;
898 c->right_mb = ( bounds->x + bounds->w ) / c->mb_w - 1;
899 c->bottom_mb = ( bounds->y + bounds->h ) / c->mb_h - 1;
900 c->bounds.x = bounds->x;
901 c->bounds.y = bounds->y;
902 c->bounds.w = bounds->w;
903 c->bounds.h = bounds->h;
905 c->left_mb = c->prev_left_mb = 0;
906 c->top_mb = c->prev_top_mb = 0;
907 c->right_mb = c->prev_right_mb = c->mv_buffer_width - 1; // Zero indexed
908 c->bottom_mb = c->prev_bottom_mb = c->mv_buffer_height - 1;
911 c->bounds.w = *width;
912 c->bounds.h = *height;
915 // If video is advancing, run motion vector algorithm and etc...
916 if( c->former_frame_position + 1 == c->current_frame_position )
919 // Swap the motion vector buffers and reuse allocated memory
920 struct motion_vector_s *temp = c->current_vectors;
921 c->current_vectors = c->former_vectors;
922 c->former_vectors = temp;
924 // This is done because filter_vismv doesn't pay attention to frame boundry
925 memset( c->current_vectors, 0, c->mv_size );
927 // Perform the motion search
928 motion_search( c->cache_image, *image, c );
930 collect_post_statistics( c );
933 // Detect shot changes
934 if( c->comparison_average > 10 * c->mb_w * c->mb_h &&
935 c->comparison_average > c->previous_msad * 2 )
937 fprintf(stderr, " - SAD: %d <<Shot change>>\n", c->comparison_average);
938 mlt_properties_set_int( MLT_FRAME_PROPERTIES( frame ), "shot_change", 1);
939 // c->former_vectors_valid = 0; // Invalidate the previous frame's predictors
943 c->former_vectors_valid = 1;
945 //fprintf(stderr, " - SAD: %d\n", c->comparison_average);
948 c->previous_msad = c->comparison_average;
950 if( c->comparison_average != 0 ) { // If the frame is not a duplicate of the previous frame
952 // denoise the vector buffer
954 median_denoise( c->current_vectors, c );
956 // Pass the new vector data into the frame
957 mlt_properties_set_data( MLT_FRAME_PROPERTIES( frame ), "motion_est.vectors",
958 (void*)c->current_vectors, c->mv_size, NULL, NULL );
960 // Cache the frame's image. Save the old cache. Reuse memory.
961 // After this block, exactly two unique frames will be cached
962 uint8_t *timg = c->cache_image;
963 c->cache_image = c->former_image;
964 c->former_image = timg;
965 memcpy( c->cache_image, *image, *width * *height * c->xstride );
970 // Undo the Swap, This fixes the ugliness caused by a duplicate frame
971 temp = c->current_vectors;
972 c->current_vectors = c->former_vectors;
973 c->former_vectors = temp;
974 mlt_properties_set_data( MLT_FRAME_PROPERTIES( frame ), "motion_est.vectors",
975 (void*)c->former_vectors, c->mv_size, NULL, NULL );
979 if( c->shot_change == 1)
981 else if( c->show_reconstruction )
982 show_reconstruction( *image, c );
983 else if( c->show_residual )
984 show_residual( *image, c );
988 else if( c->former_frame_position == c->current_frame_position )
990 // Pass the old vector data into the frame if it's valid
991 if( c->former_vectors_valid == 1 ) {
992 mlt_properties_set_data( MLT_FRAME_PROPERTIES( frame ), "motion_est.vectors",
993 (void*)c->current_vectors, c->mv_size, NULL, NULL );
995 if( c->shot_change == 1)
997 else if( c->toggle_when_paused == 1 ) {
998 if( c->show_reconstruction )
999 show_reconstruction( *image, c );
1000 else if( c->show_residual )
1001 show_residual( *image, c );
1002 c->toggle_when_paused = 2;
1004 else if( c->toggle_when_paused == 2 )
1005 c->toggle_when_paused = 1;
1007 if( c->show_reconstruction )
1008 show_reconstruction( *image, c );
1009 else if( c->show_residual )
1010 show_residual( *image, c );
1015 mlt_properties_set_int( MLT_FRAME_PROPERTIES( frame ), "shot_change", c->shot_change);
1017 // there was jump in frame number
1019 // fprintf(stderr, "Warning: there was a frame number jumped from %d to %d.\n", c->former_frame_position, c->current_frame_position);
1020 c->former_vectors_valid = 0;
1024 // Cache our bounding geometry for the next frame's processing
1025 c->prev_left_mb = c->left_mb;
1026 c->prev_top_mb = c->top_mb;
1027 c->prev_right_mb = c->right_mb;
1028 c->prev_bottom_mb = c->bottom_mb;
1030 // Remember which frame this is
1031 c->former_frame_position = c->current_frame_position;
1034 mlt_properties_set_int( MLT_FRAME_PROPERTIES( frame ), "motion_est.macroblock_width", c->mb_w );
1035 mlt_properties_set_int( MLT_FRAME_PROPERTIES( frame ), "motion_est.macroblock_height", c->mb_h );
1036 mlt_properties_set_int( MLT_FRAME_PROPERTIES( frame ), "motion_est.left_mb", c->left_mb );
1037 mlt_properties_set_int( MLT_FRAME_PROPERTIES( frame ), "motion_est.right_mb", c->right_mb );
1038 mlt_properties_set_int( MLT_FRAME_PROPERTIES( frame ), "motion_est.top_mb", c->top_mb );
1039 mlt_properties_set_int( MLT_FRAME_PROPERTIES( frame ), "motion_est.bottom_mb", c->bottom_mb );
1042 struct timeval finish; gettimeofday(&finish, NULL ); int difference = (finish.tv_sec - start.tv_sec) * 1000000 + (finish.tv_usec - start.tv_usec);
1043 fprintf(stderr, " in frame %d:%d usec\n", c->current_frame_position, difference);
1052 /** filter processing.
1055 static mlt_frame filter_process( mlt_filter this, mlt_frame frame )
1058 // Keeps tabs on the filter object
1059 mlt_frame_push_service( frame, this);
1061 // Push the frame filter
1062 mlt_frame_push_get_image( frame, filter_get_image );
1067 /** Constructor for the filter.
1069 mlt_filter filter_motion_est_init( mlt_profile profile, mlt_service_type type, const char *id, char *arg )
1071 mlt_filter this = mlt_filter_new( );
1074 // Get the properties object
1075 mlt_properties properties = MLT_FILTER_PROPERTIES( this );
1077 // Initialize the motion estimation context
1078 struct motion_est_context_s *context;
1079 context = mlt_pool_alloc( sizeof(struct motion_est_context_s) );
1080 mlt_properties_set_data( properties, "context", (void *)context, sizeof( struct motion_est_context_s ),
1081 mlt_pool_release, NULL );
1084 // Register the filter
1085 this->process = filter_process;
1087 /* defaults that may be overridden */
1090 context->skip_prediction = 0;
1091 context->limit_x = 64;
1092 context->limit_y = 64;
1093 context->search_method = DIAMOND_SEARCH; // FIXME: not used
1094 context->check_chroma = 0;
1095 context->denoise = 1;
1096 context->show_reconstruction = 0;
1097 context->show_residual = 0;
1098 context->toggle_when_paused = 0;
1100 /* reference functions that may have optimized versions */
1101 context->compare_reference = sad_reference;
1103 // The rest of the buffers will be initialized when the filter is first processed
1104 context->initialized = 0;