]> git.sesse.net Git - mlt/commitdiff
This is a significant rewrite.
authordezeroex <dezeroex@d19143bc-622f-0410-bfdd-b5b2a6649095>
Sun, 7 Aug 2005 21:54:39 +0000 (21:54 +0000)
committerdezeroex <dezeroex@d19143bc-622f-0410-bfdd-b5b2a6649095>
Sun, 7 Aug 2005 21:54:39 +0000 (21:54 +0000)
-Cleared up as many conceptualy sticky points as possible.
-Removed chroma comparison code pending a better rewrite.
-Added show_residual=1 and show_reconstruction=1 debug modes. See README.
-Renamed many variables and functions.
-Revamped geometry handling.
-Lots more I'm forgeting.

git-svn-id: https://mlt.svn.sourceforge.net/svnroot/mlt/trunk/mlt@797 d19143bc-622f-0410-bfdd-b5b2a6649095

src/modules/motion_est/filter_motion_est.c

index 6fa38ed58e39d655a6632b10dce02f410f6e542b..1b0058d3867e0c975001fe67431ba34f0b30e1be 100644 (file)
@@ -6,7 +6,7 @@
  *     Vector optimization coming soon. 
  *
  *     Vector orientation: The vector data that is generated for the current frame specifies
- *     the motion from the previous frame to the current frame. Thus, to know how a macroblock
+ *     the motion from the previous frame to the current frame. To know how a macroblock
  *     in the current frame will move in the future, the next frame is needed.
  *
  * This program is free software; you can redistribute it and/or modify
 #include <math.h>
 #include <string.h>
 #include <sys/time.h>
-#include <assert.h>
+#include <unistd.h>
 
 #include "sad_sse.h"
 
+#define NDEBUG
+#include <assert.h>
 
 #undef DEBUG
 #undef DEBUG_ASM
 #define MIN(a,b) ((a) > (b) ? (b) : (a))
 #define ABS(a) ((a) >= 0 ? (a) : (-(a)))
 
-#ifdef COUNT_COMPARES
-       int compares;
-#endif
 
 typedef struct motion_vector_s motion_vector;
 
-struct yuv_data
-{
-       uint8_t *y;
-       uint8_t *u;
-       uint8_t *v;
-
-};
-
 struct motion_est_context_s
 {
-       int initialized;                                //<! true if filter has been initialized
+       int initialized;                                // true if filter has been initialized
+
+#ifdef COUNT_COMPARES
+       int compares;
+#endif
 
        /* same as mlt_frame's parameters */
        int width, height;
 
        /* Operational details */
-       int macroblock_width, macroblock_height;
+       int mb_w, mb_h;
        int xstride, ystride;
-       //uint8_t *former_image;                        //<! Copy of the previous frame's image
-       struct yuv_data former_image, current_image;
-       int search_method, skip_prediction, shot_change;
-       int limit_x, limit_y;                   //<! max x and y of a motion vector
-       int edge_blocks_x, edge_blocks_y;
+       uint8_t *cache_image;                   // Copy of current frame
+       uint8_t *former_image;                  // Copy of former frame 
+       int search_method;
+       int skip_prediction;
+       int shot_change;
+       int limit_x, limit_y;                   // max x and y of a motion vector
        int initial_thresh;
        int check_chroma;                       // if check_chroma == 1 then compare chroma
        int denoise;
        int previous_msad;
+       int show_reconstruction;
+       int toggle_when_paused;
+       int show_residual;
 
        /* bounds */
-       struct mlt_geometry_item_s prev_bounds; // Cache last frame's bounds (needed for predictor vectors validity)
-       struct mlt_geometry_item_s *bounds;     // Current bounds
+       struct mlt_geometry_item_s bounds;      // Current bounds (from filters crop_detect, autotrack rectangle, or other)
 
-       /* bounds in macroblock units */
+       /* bounds in macroblock units; macroblocks are completely contained within the boundry */
        int left_mb, prev_left_mb, right_mb, prev_right_mb;
        int top_mb, prev_top_mb, bottom_mb, prev_bottom_mb;
 
@@ -95,12 +93,13 @@ struct motion_est_context_s
 
        /* vector buffers */
        int former_vectors_valid;               //<! true if the previous frame's buffered motion vector data is valid
-       motion_vector *former_vectors, *current_vectors;
+       motion_vector *former_vectors;
+       motion_vector *current_vectors;
        motion_vector *denoise_vectors;
        mlt_position former_frame_position, current_frame_position;
 
-       /* two metrics for diagnostics. lower is a better estimation but beware of local minima  */
-       float predictive_misses;                // How often do the prediction metion vectors fail?
+       /* diagnostic metrics */
+       float predictive_misses;                // How often do the prediction motion vectors fail?
        int comparison_average;                 // How far does the best estimation deviate from a perfect comparison?
        int bad_comparisons;
        int average_length;
@@ -112,60 +111,56 @@ struct motion_est_context_s
 
 };
 
-
-// Clip the macroblocks as required. Only used for blocks at the edge of the picture
-// "from" is assumed to be unclipped
-inline static int clip( int *from_x,
-                       int *from_y,
-                       int *to_x,
-                       int *to_y,
-                       int *w,                 //<! macroblock width
-                       int *h,                 //<! macroblock height
-                       int width,              //<! image width
-                       int height)             //<! image height
+// This is used to constrains pixel operations between two blocks to be within the image boundry
+inline static int constrain(   int *x, int *y, int *w, int *h,                 
+                               const int dx, const int dy,
+                               const int left, const int right,
+                               const int top, const int bottom)
 {
-
-       uint32_t penalty = 1 << SHIFT;  // Retain a few extra bits of precision minus floating-point's blemishes
-       int diff;
-
-       // Origin of macroblock moves left of absolute boundy
-       if( *to_x < 0 ) {
-               if( *to_x + *w <= 0) return 0;                  // Clipped out of existance
-               penalty = (*w * penalty) / (*w + *to_x);         // Recipricol of the fraction of the block that remains
-               *from_x -= *to_x;
-               *w += *to_x; 
-               *to_x = 0;
-       }
-       // Portion of macroblock moves right of absolute boundry
-       else if( *to_x + *w > width ) {
-               if(*to_x >= width) return 0;                    // Clipped out of existance
-               diff  = *to_x + *w - width;                     // Width of area clipped (0 < diff <  macroblock width)
-               penalty = (*w * penalty) / (*w - diff);         // Recipricol of the fraction of the block that remains
-               *w -= diff;
+       uint32_t penalty = 1 << SHIFT;                  // Retain a few extra bits of precision
+       int x2 = *x + dx;
+       int y2 = *y + dy;
+       int w_remains = *w;
+       int h_remains = *h;
+
+       // Origin of macroblock moves left of image boundy
+       if( *x < left || x2 < left ) {
+               w_remains = *w - left + ((*x < x2) ?  *x : x2);
+               *x += *w - w_remains;
        }
-       // Origin of macroblock moves above absolute boundy
-       if( *to_y < 0 ) {
-               if( *to_y + *h <= 0) return 0;                  // Clipped out of existance
-               penalty = (*h * penalty) / (*h + *to_y);        // Recipricol of the fraction of the block that remains
-               *from_y -= *to_y;
-               *h += *to_y;
-               *to_y = 0;
-       }
-       // Portion of macroblock moves bellow absolute boundry
-       else if( *to_y + *h > height ) {
-               if(*to_y >= height) return 0;                   // Clipped out of existance
-               diff = *to_y + *h - height;                     // Height of area clipped (0 < diff < macroblock height)
-               penalty = (*h * penalty) / (*h - diff);         // Recipricol of the fraction of the block that is clipped
-               *h -= diff;
+       // Portion of macroblock moves right of image boundry
+       else if( *x + *w > right || x2 + *w > right )
+               w_remains = right - ((*x > x2) ? *x : x2);
+
+       // Origin of macroblock moves above image boundy
+       if( *y < top || y2 < top ) {
+               h_remains = *h - top + ((*y < y2) ? *y : y2);
+               *y += *h - h_remains;
        }
+       // Portion of macroblock moves bellow image boundry
+       else if( *y + *h > bottom || y2 + *h > bottom )
+               h_remains = bottom - ((*y > y2) ?  *y : y2);
+
+       if( w_remains == *w && h_remains == *h ) return penalty; 
+       if( w_remains <= 0 || h_remains <= 0) return 0; // Block is clipped out of existance
+       penalty = (*w * *h * penalty)
+               / ( w_remains * h_remains);             // Recipricol of the fraction of the block that remains
+
+       assert(*x >= left); assert(x2 + *w - w_remains >= left);
+       assert(*y >= top); assert(y2 + *h - h_remains >= top);
+       assert(*x + w_remains <= right); assert(x2 + w_remains <= right);
+       assert(*y + h_remains <= bottom); assert(y2 + h_remains <= bottom);
+
+       *w = w_remains;                                 // Update the width and height
+       *h = h_remains;
+
        return penalty;
 }
 
-
 /** /brief Reference Sum of Absolute Differences comparison function
 *
 */
-inline static int sad_reference( uint8_t *block1, uint8_t *block2, int xstride, int ystride, int w, int h )
+inline static int sad_reference( uint8_t *block1, uint8_t *block2, const int xstride, const int ystride, const int w, const int h )
 {
        int i, j, score = 0;
        for ( j = 0; j < h; j++ ){
@@ -179,98 +174,69 @@ inline static int sad_reference( uint8_t *block1, uint8_t *block2, int xstride,
        return score;
 }
 
-inline static void change_422_to_444_planar_rep( uint8_t *image, struct yuv_data yuv, struct motion_est_context_s *c )
-{
-       register uint8_t *p = image;
-       register uint8_t *q = image + c->width * c->height * 2;
-       while ( *p != *q ) {
-               *(yuv.y++) = *(p ++);
-               *(yuv.u++) = *p;
-               *(yuv.u++) = *(p ++);
-               *(yuv.y++) = *(p ++);
-               *(yuv.v++) = *p;
-               *(yuv.v++) = *(p ++);
-       }
-}
-
-// broken
-inline static void change_420p_to_444_planar_rep( uint8_t *image, struct yuv_data yuv, struct motion_est_context_s *c )
-{
-       uint8_t *p = image + c->width * c->height;
-       uint8_t *q = p + c->width*c->height/2;
-       uint8_t *u2, *v2;
-       while( *p != *q ) {
-               u2 = yuv.u + c->width;
-               *yuv.u  ++ = *p;
-               *yuv.u  ++ = *p;
-               *u2 ++ = *p;
-               *u2 ++ = *p ++;
-       }
-
-       *q += c->width*c->height/2;
-       while( *p != *q ) {
-               v2 = yuv.v + c->width;
-               *yuv.v  ++ = *p;
-               *yuv.v  ++ = *p;
-               *v2 ++ = *p;
-               *v2 ++ = *p ++;
-       }
-
-}
 
 /** /brief Abstracted block comparison function
 */
-inline static int compare( uint8_t *from,
-                          uint8_t *to,
-                          int from_x,
-                          int from_y,
-                          int to_x,
-                          int to_y,
+inline static int block_compare( uint8_t *block1,
+                          uint8_t *block2,
+                          int x,
+                          int y,
+                          int dx,
+                          int dy,
                           struct motion_est_context_s *c)
 {
+
 #ifdef COUNT_COMPARES
-       compares++;
+       c->compares++;
 #endif
 
-       if( ABS(from_x - to_x) >= c->limit_x || ABS(from_y - to_y) >= c->limit_y )
-               return MAX_MSAD;
-
        int score;
+
+       // Default comparison may be overridden by the slower, more capable reference comparison 
        int (*cmp)(uint8_t *, uint8_t *, int, int, int, int) = c->compare_optimized;
 
-       int mb_w = c->macroblock_width;
-       int mb_h = c->macroblock_height;
+       // vector displacement limited has been exceeded
+       if( ABS( dx ) >= c->limit_x || ABS( dy ) >= c->limit_y )
+               return MAX_MSAD;
+
+       int mb_w = c->mb_w;     // Some writeable local copies
+       int mb_h = c->mb_h;
 
-       int penalty = clip(&from_x, &from_y, &to_x, &to_y, &mb_w, &mb_h, c->width, c->height);
-       if ( penalty == 1<<SHIFT)
-           penalty = clip(&to_x, &to_y, &from_x, &from_y, &mb_w, &mb_h, c->width, c->height);
+       // Determine if either macroblock got clipped
+       int penalty = constrain( &x, &y, &mb_w, &mb_h, dx, dy, 0, c->width, 0, c->height);
 
-       if( penalty == 0 )                      // Clipped out of existance
+       // Some gotchas
+       if( penalty == 0 )                      // Clipped out of existance: Return worst score
                return MAX_MSAD;
-       else if( penalty != 1<<SHIFT )          // SIMD optimized comparison won't work
+       else if( penalty != 1<<SHIFT )          // Nonstandard macroblock dimensions: Disable SIMD optimizizations. 
                cmp = c->compare_reference;
 
-       uint8_t *from_block = from + from_x * c->xstride + from_y * c->ystride; 
-       uint8_t *to_block = to + to_x * c->xstride + to_y * c->ystride; 
+       // Calculate the memory locations of the macroblocks
+       block1 += x      * c->xstride + y       * c->ystride; 
+       block2 += (x+dx) * c->xstride + (y+dy)  * c->ystride; 
 
        #ifdef DEBUG_ASM        
        if( penalty == 1<<SHIFT ){
-               score = c->compare_reference( from_block, to_block, c->xstride, c->ystride, mb_w, mb_h );
-               int score2 = c->compare_optimized( from_block, to_block, c->xstride, c->ystride, mb_w, mb_h );
+               score = c->compare_reference( block1, block2, c->xstride, c->ystride, mb_w, mb_h );
+               int score2 = c->compare_optimized( block1, block2, c->xstride, c->ystride, mb_w, mb_h );
                if ( score != score2 )
                        fprintf(stderr, "Your assembly doesn't work! Reference: %d Asm: %d\n", score, score2);
        }
        else
        #endif
 
-       score = cmp( from_block, to_block, c->xstride, c->ystride, mb_w, mb_h );
+       score = cmp( block1, block2, c->xstride, c->ystride, mb_w, mb_h );
 
-       return ( score * penalty ) >> SHIFT;                    // The extra precision is no longer wanted
+       return ( score * penalty ) >> SHIFT;                    // Ditch the extra precision
 }
 
-static inline void check_candidates (   struct yuv_data *from, struct yuv_data *to,
-                                       int from_x, int from_y,
-                                       motion_vector *candidates, int count, int unique,
+static inline void check_candidates (   uint8_t *ref,
+                                       uint8_t *candidate_base,
+                                       const int x,
+                                       const int y,
+                                       const motion_vector *candidates,// Contains to_x & to_y
+                                       const int count,                // Number of candidates
+                                       const int unique,               // Sometimes we know the candidates are unique
                                        motion_vector *result,
                                        struct motion_est_context_s *c )
 {
@@ -292,30 +258,11 @@ static inline void check_candidates (   struct yuv_data *from, struct yuv_data *
                        }
 
                        // Luma
-                       score = compare( from->y, to->y, from_x, from_y,
-                                        from_x + candidates[i].dx,     /* to x */
-                                        from_y + candidates[i].dy,     /* to y */
-                                        c);
-
-                       if ( c->check_chroma ) {
-                               if ( score >= result->msad )    // Early term
-                                       continue;
-
-                               // Chroma - U
-                               score += compare( from->u, to->u, from_x, from_y,
-                                                from_x + candidates[i].dx,     /* to x */
-                                                from_y + candidates[i].dy,     /* to y */
-                                                c);
-       
-                               if ( score >= result->msad )    // Early term
-                                       continue;
-       
-                               // Chroma - V
-                               score += compare( from->v, to->v, from_x, from_y,
-                                                from_x + candidates[i].dx,     /* to x */
-                                                from_y + candidates[i].dy,     /* to y */
-                                                c);
-                       }
+                       score = block_compare( ref, candidate_base,
+                                               x, y,
+                                               candidates[i].dx,       // from
+                                               candidates[i].dy,
+                                               c);
 
                        if ( score < result->msad ) {   // New minimum
                                result->dx = candidates[i].dx;
@@ -330,12 +277,12 @@ static inline void check_candidates (   struct yuv_data *from, struct yuv_data *
 * Operates on a single macroblock
 */
 static inline void diamond_search(
-                       struct yuv_data *from,                  //<! Image data from previous frame
-                       struct yuv_data *to,                    //<! Image data in current frame
-                       int mb_x,                       //<! X upper left corner of macroblock
-                       int mb_y,                       //<! U upper left corner of macroblock
-                       struct motion_vector_s *result, //<! Best predicted mv and eventual result
-                       struct motion_est_context_s *c) //<! motion estimation context
+                       uint8_t *ref,                           //<! Image data from previous frame
+                       uint8_t *candidate_base,                //<! Image data in current frame
+                       const int x,                            //<! X upper left corner of macroblock
+                       const int y,                            //<! U upper left corner of macroblock
+                       struct motion_vector_s *result,         //<! Best predicted mv and eventual result
+                       struct motion_est_context_s *c)         //<! motion estimation context
 {
 
        // diamond search pattern
@@ -382,44 +329,48 @@ static inline void diamond_search(
                        former.dx = best.dx; former.dy = best.dy;       // Keep track of new former best
                }
        
-               check_candidates ( from, to, mb_x, mb_y, candidates, i, 1, result, c ); 
+               check_candidates ( ref, candidate_base, x, y, candidates, i, 1, result, c ); 
+
+               // Which candidate was the best?
                best.dx = result->dx - current.dx;
                best.dy = result->dy - current.dy;
 
+               // A better canidate was not found
                if ( best.dx == 0 && best.dy == 0 )
                        return;
 
                if ( first == 1 ){
                        first = 0;
-                       former.dx = best.dx; former.dy = best.dy; // First iteration, sensible value for former_d*
+                       former.dx = best.dx; former.dy = best.dy; // First iteration, sensible value for former.d*
                }
        }
 }
 
 /* /brief Full (brute) search 
 * Operates on a single macroblock
-*/
+*/     
+__attribute__((used))
 static void full_search(
-                       struct yuv_data *from,                  //<! Image data from previous frame
-                       struct yuv_data *to,                    //<! Image data in current frame
-                       int mb_x,                       //<! X upper left corner of macroblock
-                       int mb_y,                       //<! U upper left corner of macroblock
-                       struct motion_vector_s *result, //<! Best predicted mv and eventual result
-                       struct motion_est_context_s *c) //<! motion estimation context
+                       uint8_t *ref,                           //<! Image data from previous frame
+                       uint8_t *candidate_base,                //<! Image data in current frame
+                       int x,                                  //<! X upper left corner of macroblock
+                       int y,                                  //<! U upper left corner of macroblock
+                       struct motion_vector_s *result,         //<! Best predicted mv and eventual result
+                       struct motion_est_context_s *c)         //<! motion estimation context
 {
        // Keep track of best candidate
        int i,j,score;
 
        // Go loopy
-       for( i = -c->macroblock_width; i <= c->macroblock_width; i++ ){
-               for( j = -c->macroblock_height; j <=  c->macroblock_height; j++ ){
+       for( i = -c->mb_w; i <= c->mb_w; i++ ){
+               for( j = -c->mb_h; j <=  c->mb_h; j++ ){
 
-                       score = compare( from->y, to->y,
-                                        mb_x,          /* from x */
-                                        mb_y,          /* from y */
-                                        mb_x + i,      /* to x */
-                                        mb_y + j,      /* to y */
-                                        c);            /* context */
+                       score = block_compare( ref, candidate_base,
+                                               x,
+                                               y,
+                                               x + i,
+                                               y + j,
+                                               c);
 
                        if ( score < result->msad ) {
                                result->dx = i;
@@ -430,23 +381,6 @@ static void full_search(
        }
 }
 
-// Credits: ffmpeg
-// return the median
-static inline int median_predictor(int a, int b, int c) {
-       if ( a > b ){
-               if ( c > b ){
-                       if ( c > a  ) b = a;
-                       else      b = c;
-               }
-       } else {
-               if ( b > c ){
-                       if ( c > a ) b = c;
-                       else     b = a;
-               }
-       }
-       return b;
-}
-
 // Macros for pointer calculations
 #define CURRENT(i,j)   ( c->current_vectors + (j)*c->mv_buffer_width + (i) )
 #define FORMER(i,j)    ( c->former_vectors + (j)*c->mv_buffer_width + (i) )
@@ -465,8 +399,8 @@ static void median_denoise( motion_vector *v, struct motion_est_context_s *c )
        int xvalues[9], yvalues[9];
 
        int i,j,n;
-       for( i = c->left_mb; i <= c->right_mb; i++ ){
-               for( j = c->top_mb; j <= c->bottom_mb; j++ )
+       for( j = c->top_mb; j <= c->bottom_mb; j++ )
+               for( i = c->left_mb; i <= c->right_mb; i++ ){
                {
                        n = 0;
 
@@ -536,17 +470,36 @@ static void median_denoise( motion_vector *v, struct motion_est_context_s *c )
 
 }
 
+// Credits: ffmpeg
+// return the median
+static inline int median_predictor(int a, int b, int c) {
+       if ( a > b ){
+               if ( c > b ){
+                       if ( c > a  ) b = a;
+                       else      b = c;
+               }
+       } else {
+               if ( b > c ){
+                       if ( c > a ) b = c;
+                       else     b = a;
+               }
+       }
+       return b;
+}
+
+
 /** /brief Motion search
 *
+* For each macroblock in the current frame, estimate the block from the last frame that
+* matches best.
 *
-* Search for the Vector that best represents the motion *from the last frame *to the current frame
 * Vocab: Colocated - the pixel in the previous frame at the current position
 *
 * Based on enhanced predictive zonal search. [Tourapis 2002]
 */
-static void search( struct yuv_data from,                      //<! Image data. Motion vector source in previous frame
-                   struct yuv_data to,                 //<! Image data. Motion vector destination current
-                   struct motion_est_context_s *c)     //<! The context
+static void motion_search( uint8_t *from,                      //<! Image data.
+                          uint8_t *to,                         //<! Image data. Rigid grid.
+                          struct motion_est_context_s *c)      //<! The context
 {
 
 #ifdef COUNT_COMPARES
@@ -638,17 +591,22 @@ static void search( struct yuv_data from,                 //<! Image data. Motion vector sourc
                candidates[n  ].dx = 0;
                candidates[n++].dy = 0;
 
-               int from_x = i * c->macroblock_width;
-               int from_y = j * c->macroblock_height;
-               check_candidates ( &from, &to, from_x, from_y, candidates, n, 0, here, c ); 
+               int x = i * c->mb_w;
+               int y = j * c->mb_h;
+               check_candidates ( to, from, x, y, candidates, n, 0, here, c ); 
 
 
 #ifndef FULLSEARCH
-               diamond_search( &from, &to, from_x, from_y, here, c); 
+               diamond_search( to, from, x, y, here, c); 
 #else
-               full_search( from, to, from_x, from_y, here, c); 
+               full_search( to, from, x, y, here, c); 
 #endif
 
+               assert( x + c->mb_w + here->dx > 0 );   // All macroblocks must have area > 0
+               assert( y + c->mb_h + here->dy > 0 );
+               assert( x + here->dx < c->width );
+               assert( y + here->dy < c->height );
+
         } /* End column loop */
        } /* End row loop */
 
@@ -693,46 +651,123 @@ void collect_post_statistics( struct motion_est_context_s *c ) {
 
 static void init_optimizations( struct motion_est_context_s *c )
 {
-       if ( c->check_chroma ) {
-               switch(c->macroblock_width){
-                       case 8:  if(c->macroblock_height == 8)  c->compare_optimized = sad_sse_8x8;
-                                else                           c->compare_optimized = sad_sse_8w;
-                                break;
-                       case 16: if(c->macroblock_height == 16) c->compare_optimized = sad_sse_16x16;
-                                else                           c->compare_optimized = sad_sse_16w;
-                                break;
-                       case 32: if(c->macroblock_height == 32) c->compare_optimized = sad_sse_32x32;
-                                else                           c->compare_optimized = sad_sse_32w;
-                                break;
-                       case 64: c->compare_optimized = sad_sse_64w;
-                                break;
-                       default: c->compare_optimized = sad_reference;
-                                break;
-               }
+       switch(c->mb_w){
+               case 4:  if(c->mb_h == 4)       c->compare_optimized = sad_sse_422_luma_4x4;
+                        else                           c->compare_optimized = sad_sse_422_luma_4w;
+                        break;
+               case 8:  if(c->mb_h == 8)  c->compare_optimized = sad_sse_422_luma_8x8;
+                        else                           c->compare_optimized = sad_sse_422_luma_8w;
+                        break;
+               case 16: if(c->mb_h == 16) c->compare_optimized = sad_sse_422_luma_16x16;
+                        else                           c->compare_optimized = sad_sse_422_luma_16w;
+                        break;
+               case 32: if(c->mb_h == 32) c->compare_optimized = sad_sse_422_luma_32x32;
+                        else                           c->compare_optimized = sad_sse_422_luma_32w;
+                        break;
+               case 64: c->compare_optimized = sad_sse_422_luma_64w;
+                        break;
+               default: c->compare_optimized = sad_reference;
+                        break;
        }
-       else
+}
+
+inline static void set_red(uint8_t *image, struct motion_est_context_s *c)
+{
+       int n;
+       for( n = 0; n < c->width * c->height * 2; n+=4 )
        {
-               switch(c->macroblock_width){
-                       case 4:  if(c->macroblock_height == 4)  c->compare_optimized = sad_sse_422_luma_4x4;
-                                else                           c->compare_optimized = sad_sse_422_luma_4w;
-                                break;
-                       case 8:  if(c->macroblock_height == 8)  c->compare_optimized = sad_sse_422_luma_8x8;
-                                else                           c->compare_optimized = sad_sse_422_luma_8w;
-                                break;
-                       case 16: if(c->macroblock_height == 16) c->compare_optimized = sad_sse_422_luma_16x16;
-                                else                           c->compare_optimized = sad_sse_422_luma_16w;
-                                break;
-                       case 32: if(c->macroblock_height == 32) c->compare_optimized = sad_sse_422_luma_32x32;
-                                else                           c->compare_optimized = sad_sse_422_luma_32w;
-                                break;
-                       case 64: c->compare_optimized = sad_sse_422_luma_64w;
-                                break;
-                       default: c->compare_optimized = sad_reference;
-                                break;
+               image[n]   = 79;
+               image[n+1] = 91;
+               image[n+2] = 79;
+               image[n+3] = 237;
+       }
+               
+}
+
+static void show_residual( uint8_t *result,  struct motion_est_context_s *c )
+{
+       int i, j;
+       int x,y,w,h;
+       int dx, dy;
+       int tx,ty;
+       uint8_t *b, *r;
+
+//     set_red(result,c);
+
+       for( j = c->top_mb; j <= c->bottom_mb; j++ ){  
+        for( i = c->left_mb; i <= c->right_mb; i++ ){
+
+               dx = CURRENT(i,j)->dx;
+               dy = CURRENT(i,j)->dy;
+               w = c->mb_w;
+               h = c->mb_h;
+               x = i * w;
+               y = j * h;
+
+               // Denoise function caused some blocks to be completely clipped, ignore them    
+               if (constrain( &x, &y, &w, &h, dx, dy, 0, c->width, 0, c->height) == 0 )
+                       continue;
+
+               for( ty = y; ty < y + h ; ty++ ){
+                for( tx = x; tx < x + w ; tx++ ){
+
+                       b = c->former_image +  (tx+dx)*c->xstride + (ty+dy)*c->ystride;
+                       r = result +            tx*c->xstride + ty*c->ystride;
+
+                       r[0] = 16 + ABS( r[0] - b[0] );
+
+                       if( dx % 2 == 0 )
+                               r[1] = 128 + ABS( r[1] - b[1] );
+                       else
+                               // FIXME: may exceed boundies
+                               r[1] = 128 + ABS( r[1] - ( *(b-1) + b[3] ) /2 );
+                }
                }
+        }
        }
 }
-       
+
+static void show_reconstruction( uint8_t *result, struct motion_est_context_s *c )
+{
+       int i, j;
+       int x,y,w,h;
+       int dx,dy;
+       uint8_t *r, *s;
+       int tx,ty;
+
+       for( i = c->left_mb; i <= c->right_mb; i++ ){
+        for( j = c->top_mb; j <= c->bottom_mb; j++ ){  
+
+               dx = CURRENT(i,j)->dx;
+               dy = CURRENT(i,j)->dy;
+               w = c->mb_w;
+               h = c->mb_h;
+               x = i * w;
+               y = j * h;
+
+               // Denoise function caused some blocks to be completely clipped, ignore them    
+               if (constrain( &x, &y, &w, &h, dx, dy, 0, c->width, 0, c->height) == 0 )
+                       continue;
+
+               for( ty = y; ty < y + h ; ty++ ){
+                for( tx = x; tx < x + w ; tx++ ){
+
+                       r = result +          tx*c->xstride + ty*c->ystride;
+                       s = c->former_image + (tx+dx)*c->xstride + (ty+dy)*c->ystride;
+
+                       r[0] = s[0];
+
+                       if( dx % 2 == 0 )
+                               r[1] = s[1];
+                       else
+                               // FIXME: may exceed boundies
+                               r[1] = ( *(s-1) + s[3] ) /2;
+                }
+               }
+        }
+       }
+}
+
 // Image stack(able) method
 static int filter_get_image( mlt_frame frame, uint8_t **image, mlt_image_format *format, int *width, int *height, int writable )
 {
@@ -742,17 +777,14 @@ static int filter_get_image( mlt_frame frame, uint8_t **image, mlt_image_format
        // Get the motion_est context object
        struct motion_est_context_s *c = mlt_properties_get_data( MLT_FILTER_PROPERTIES( filter ), "context", NULL);
 
-               #ifdef BENCHMARK
-               struct timeval start; gettimeofday(&start, NULL );
-               #endif
 
        // Get the new image and frame number
        int error = mlt_frame_get_image( frame, image, format, width, height, 1 );
 
-               #ifdef BENCHMARK
-               struct timeval finish; gettimeofday(&finish, NULL ); int difference = (finish.tv_sec - start.tv_sec) * 1000000 + (finish.tv_usec - start.tv_usec);
-               fprintf(stderr, " in frame %d:%d usec\n", c->current_frame_position, difference);
-               #endif
+       #ifdef BENCHMARK
+       struct timeval start; gettimeofday(&start, NULL );
+       #endif
+
 
        if( error != 0 )
                mlt_properties_debug( MLT_FRAME_PROPERTIES(frame), "error after mlt_frame_get_image() in motion_est", stderr );
@@ -770,15 +802,15 @@ static int filter_get_image( mlt_frame frame, uint8_t **image, mlt_image_format
 
                /* Get parameters that may have been overridden */
                if( mlt_properties_get( properties, "macroblock_width") != NULL )
-                       c->macroblock_width = mlt_properties_get_int( properties, "macroblock_width");
+                       c->mb_w = mlt_properties_get_int( properties, "macroblock_width");
 
                if( mlt_properties_get( properties, "macroblock_height") != NULL )
-                       c->macroblock_height = mlt_properties_get_int( properties, "macroblock_height");
+                       c->mb_h = mlt_properties_get_int( properties, "macroblock_height");
 
                if( mlt_properties_get( properties, "prediction_thresh") != NULL )
                        c->initial_thresh = mlt_properties_get_int( properties, "prediction_thresh" );
                else
-                       c->initial_thresh = c->macroblock_width * c->macroblock_height;
+                       c->initial_thresh = c->mb_w * c->mb_h;
 
                if( mlt_properties_get( properties, "search_method") != NULL )
                        c->search_method = mlt_properties_get_int( properties, "search_method");
@@ -798,11 +830,20 @@ static int filter_get_image( mlt_frame frame, uint8_t **image, mlt_image_format
                if( mlt_properties_get( properties, "denoise" ) != NULL )
                        c->denoise = mlt_properties_get_int( properties, "denoise" );
 
+               if( mlt_properties_get( properties, "show_reconstruction" ) != NULL )
+                       c->show_reconstruction = mlt_properties_get_int( properties, "show_reconstruction" );
+
+               if( mlt_properties_get( properties, "show_residual" ) != NULL )
+                       c->show_residual = mlt_properties_get_int( properties, "show_residual" );
+
+               if( mlt_properties_get( properties, "toggle_when_paused" ) != NULL )
+                       c->toggle_when_paused = mlt_properties_get_int( properties, "toggle_when_paused" );
+
                init_optimizations( c );
 
                // Calculate the dimensions in macroblock units
-               c->mv_buffer_width = (*width / c->macroblock_width);
-               c->mv_buffer_height = (*height / c->macroblock_height);
+               c->mv_buffer_width = (*width / c->mb_w);
+               c->mv_buffer_height = (*height / c->mb_h);
 
                // Size of the motion vector buffer
                c->mv_size =  c->mv_buffer_width * c->mv_buffer_height * sizeof(struct motion_vector_s);
@@ -817,51 +858,28 @@ static int filter_get_image( mlt_frame frame, uint8_t **image, mlt_image_format
                mlt_properties_set_data( properties, "former_motion_vectors", (void *)c->former_vectors, 0, mlt_pool_release, NULL );
                mlt_properties_set_data( properties, "denoise_motion_vectors", (void *)c->denoise_vectors, 0, mlt_pool_release, NULL );
 
-
                c->former_vectors_valid = 0;
                memset( c->former_vectors, 0, c->mv_size );
 
-               // Figure out how many blocks should be considered edge blocks
-               c->edge_blocks_x = (c->limit_x + c->macroblock_width - 1) / c->macroblock_width;
-               c->edge_blocks_y = (c->limit_y + c->macroblock_height - 1) / c->macroblock_height;
-
                // Calculate the size of our steps (the number of bytes that seperate adjacent pixels in X and Y direction)
                switch( *format ) {
                        case mlt_image_yuv422:
-                               if ( c->check_chroma )
-                                       c->xstride = 1;
-                               else
-                                       c->xstride = 2;
+                               c->xstride = 2;
                                c->ystride = c->xstride * *width;
                                break; 
-/*                     case mlt_image_yuv420p:
-                               c->xstride = 1;
-                               c->ystride = c->xstride * *width;
-                               break;
-*/                     default:
+                       default:
                                // I don't know
                                fprintf(stderr, "\"I am unfamiliar with your new fangled pixel format!\" -filter_motion_est\n");
                                return -1;
                }
 
-               if ( c->check_chroma ) {
-                       // Allocate memory for the 444 images
-                       c->former_image.y = mlt_pool_alloc( *width * *height * 3 );
-                       c->current_image.y = mlt_pool_alloc( *width * *height * 3 );
-                       c->current_image.u = c->current_image.y + *width * *height;
-                       c->current_image.v = c->current_image.u + *width * *height;
-                       c->former_image.u = c->former_image.y + *width * *height;
-                       c->former_image.v = c->former_image.u + *width * *height;
-                       // Register for destruction
-                       mlt_properties_set_data( properties, "current_image", (void *)c->current_image.y, 0, mlt_pool_release, NULL );
-               }
-               else
-               {
-                       c->former_image.y = mlt_pool_alloc( *width * *height * 2 );
-               }
-               // Register for destruction
-               mlt_properties_set_data( properties, "former_image", (void *)c->former_image.y, 0, mlt_pool_release, NULL );
+               // Allocate a cache for the previous frame's image
+               c->former_image = mlt_pool_alloc( *width * *height * 2 );
+               c->cache_image = mlt_pool_alloc( *width * *height * 2 );
 
+               // Register for destruction
+               mlt_properties_set_data( properties, "cache_image", (void *)c->cache_image, 0, mlt_pool_release, NULL );
+               mlt_properties_set_data( properties, "former_image", (void *)c->former_image, 0, mlt_pool_release, NULL );
 
                c->former_frame_position = c->current_frame_position;
                c->previous_msad = 0;
@@ -870,32 +888,30 @@ static int filter_get_image( mlt_frame frame, uint8_t **image, mlt_image_format
        }
 
        /* Check to see if somebody else has given us bounds */
-       c->bounds = mlt_properties_get_data( MLT_FRAME_PROPERTIES( frame ), "bounds", NULL );
-
-       /* no bounds were given, they won't change next frame, so use a convient storage place */
-       if( c->bounds == NULL ) {
-               c->bounds = &c->prev_bounds;
-               c->bounds->x = 0;
-               c->bounds->y = 0;
-               c->bounds->w = *width - 1;      // Zero indexed
-               c->bounds->h = *height - 1;     // Zero indexed
+       struct mlt_geometry_item_s *bounds = mlt_properties_get_data( MLT_FRAME_PROPERTIES( frame ), "bounds", NULL );
+
+       if( bounds != NULL ) {
+               // translate pixel units (from bounds) to macroblock units
+               // make sure whole macroblock stays within bounds
+               c->left_mb = ( bounds->x + c->mb_w - 1 ) / c->mb_w;
+               c->top_mb = ( bounds->y + c->mb_h - 1 ) / c->mb_h;
+               c->right_mb = ( bounds->x + bounds->w ) / c->mb_w - 1;
+               c->bottom_mb = ( bounds->y + bounds->h ) / c->mb_h - 1;
+               c->bounds.x = bounds->x;
+               c->bounds.y = bounds->y;
+               c->bounds.w = bounds->w;
+               c->bounds.h = bounds->h;
+       } else {
+               c->left_mb = c->prev_left_mb = 0;
+               c->top_mb = c->prev_top_mb = 0;
+               c->right_mb = c->prev_right_mb = c->mv_buffer_width - 1;        // Zero indexed
+               c->bottom_mb = c->prev_bottom_mb = c->mv_buffer_height - 1;
+               c->bounds.x = 0;
+               c->bounds.y = 0;
+               c->bounds.w = *width;
+               c->bounds.h = *height;
        }
 
-       // translate pixel units (from bounds) to macroblock units
-       // make sure whole macroblock stays within bounds
-       c->left_mb = ( c->bounds->x + c->macroblock_width - 1 ) / c->macroblock_width;
-       c->top_mb = ( c->bounds->y + c->macroblock_height - 1 ) / c->macroblock_height;
-       c->right_mb = ( c->bounds->x + c->bounds->w ) / c->macroblock_width - 1;
-       c->bottom_mb = ( c->bounds->y + c->bounds->h ) / c->macroblock_height - 1;
-
-       // Do the same thing for the previous frame's geometry
-       // This will be used for determining validity of predictors
-       c->prev_left_mb = ( c->prev_bounds.x + c->macroblock_width - 1) / c->macroblock_width;
-       c->prev_top_mb = ( c->prev_bounds.y + c->macroblock_height - 1) / c->macroblock_height;
-       c->prev_right_mb = ( c->prev_bounds.x + c->prev_bounds.w ) / c->macroblock_width - 1;
-       c->prev_bottom_mb = ( c->prev_bounds.y + c->prev_bounds.h ) / c->macroblock_height - 1;
-
-       
        // If video is advancing, run motion vector algorithm and etc...        
        if( c->former_frame_position + 1 == c->current_frame_position )
        {
@@ -905,48 +921,17 @@ static int filter_get_image( mlt_frame frame, uint8_t **image, mlt_image_format
                c->current_vectors = c->former_vectors;
                c->former_vectors = temp;
 
-               // Swap the image buffers
-               if ( c->check_chroma ) {
-                       uint8_t *temp_yuv;
-                       temp_yuv = c->current_image.y;
-                       c->current_image.y = c->former_image.y;
-                       c->former_image.y = temp_yuv;
-                       temp_yuv = c->current_image.u;
-                       c->current_image.u = c->former_image.u;
-                       c->former_image.u = temp_yuv;
-                       temp_yuv = c->current_image.v;
-                       c->current_image.v = c->former_image.v;
-                       c->former_image.v = temp_yuv;
-
-                       switch ( *format ) {
-                               case mlt_image_yuv422:
-                                       change_422_to_444_planar_rep( *image, c->current_image, c );
-                                       break;
-                               case mlt_image_yuv420p:
-                                       change_420p_to_444_planar_rep( *image, c->current_image, c );
-                                       break;
-                               default:
-                                       break;
-                       }
-               }
-               else
-                       c->current_image.y = *image;
-
-               // Find a better place for this
+               // This is done because filter_vismv doesn't pay attention to frame boundry
                memset( c->current_vectors, 0, c->mv_size );
 
                // Perform the motion search
-
-               //collect_pre_statistics( context, *image );
-               search( c->current_image, c->former_image, c );
+               motion_search( c->cache_image, *image, c );
 
                collect_post_statistics( c );
 
 
-
-
                // Detect shot changes
-               if( c->comparison_average > 10 * c->macroblock_width * c->macroblock_height &&
+               if( c->comparison_average > 10 * c->mb_w * c->mb_h &&
                    c->comparison_average > c->previous_msad * 2 )
                {
                        fprintf(stderr, " - SAD: %d   <<Shot change>>\n", c->comparison_average);
@@ -962,7 +947,7 @@ static int filter_get_image( mlt_frame frame, uint8_t **image, mlt_image_format
 
                c->previous_msad = c->comparison_average;
 
-               if( c->comparison_average != 0 ) {
+               if( c->comparison_average != 0 ) { // If the frame is not a duplicate of the previous frame
 
                        // denoise the vector buffer
                        if( c->denoise )
@@ -972,9 +957,17 @@ static int filter_get_image( mlt_frame frame, uint8_t **image, mlt_image_format
                        mlt_properties_set_data( MLT_FRAME_PROPERTIES( frame ), "motion_est.vectors",
                                         (void*)c->current_vectors, c->mv_size, NULL, NULL );
 
+                       // Cache the frame's image. Save the old cache. Reuse memory.
+                       // After this block, exactly two unique frames will be cached
+                       uint8_t *timg = c->cache_image;
+                       c->cache_image = c->former_image;
+                       c->former_image = timg;
+                       memcpy( c->cache_image, *image, *width * *height * c->xstride );
+
+
                }
                else {
-                       // This fixes the ugliness caused by a duplicate frame
+                       // Undo the Swap, This fixes the ugliness caused by a duplicate frame
                        temp = c->current_vectors;
                        c->current_vectors = c->former_vectors;
                        c->former_vectors = temp;
@@ -982,35 +975,70 @@ static int filter_get_image( mlt_frame frame, uint8_t **image, mlt_image_format
                                         (void*)c->former_vectors, c->mv_size, NULL, NULL );
                }
 
+
+               if( c->shot_change == 1)
+                       ;
+               else if( c->show_reconstruction )
+                       show_reconstruction( *image, c );
+               else if( c->show_residual )
+                       show_residual( *image, c );
+
        }
        // paused
        else if( c->former_frame_position == c->current_frame_position )
        {
                // Pass the old vector data into the frame if it's valid
-               if( c->former_vectors_valid == 1 )
+               if( c->former_vectors_valid == 1 ) {
                        mlt_properties_set_data( MLT_FRAME_PROPERTIES( frame ), "motion_est.vectors",
                                         (void*)c->current_vectors, c->mv_size, NULL, NULL );
 
+                       if( c->shot_change == 1)
+                               ;
+                       else if( c->toggle_when_paused == 1 ) {
+                               if( c->show_reconstruction )
+                                       show_reconstruction( *image, c );
+                               else if( c->show_residual )
+                                       show_residual( *image, c );
+                               c->toggle_when_paused = 2;
+                       }
+                       else if( c->toggle_when_paused == 2 )
+                               c->toggle_when_paused = 1;
+                       else {
+                               if( c->show_reconstruction )
+                                       show_reconstruction( *image, c );
+                               else if( c->show_residual )
+                                       show_residual( *image, c );
+                       }
+
+               }
+
                mlt_properties_set_int( MLT_FRAME_PROPERTIES( frame ), "shot_change", c->shot_change);
        }
        // there was jump in frame number
-       else
+       else {
+//             fprintf(stderr, "Warning: there was a frame number jumped from %d to %d.\n", c->former_frame_position, c->current_frame_position);
                c->former_vectors_valid = 0;
+       }
 
 
        // Cache our bounding geometry for the next frame's processing
-       if( c->bounds != &c->prev_bounds )
-               memcpy( &c->prev_bounds, c->bounds, sizeof( struct mlt_geometry_item_s ) );
+       c->prev_left_mb = c->left_mb;
+       c->prev_top_mb = c->top_mb;
+       c->prev_right_mb = c->right_mb;
+       c->prev_bottom_mb = c->bottom_mb;
 
        // Remember which frame this is
        c->former_frame_position = c->current_frame_position;
 
 
-       if ( c->check_chroma == 0 )
-               memcpy( c->former_image.y, *image, *width * *height * c->xstride );
+       mlt_properties_set_int( MLT_FRAME_PROPERTIES( frame ), "motion_est.macroblock_width", c->mb_w );
+       mlt_properties_set_int( MLT_FRAME_PROPERTIES( frame ), "motion_est.macroblock_height", c->mb_h );
+
+       #ifdef BENCHMARK
+       struct timeval finish; gettimeofday(&finish, NULL ); int difference = (finish.tv_sec - start.tv_sec) * 1000000 + (finish.tv_usec - start.tv_usec);
+       fprintf(stderr, " in frame %d:%d usec\n", c->current_frame_position, difference);
+       #endif
 
-       mlt_properties_set_int( MLT_FRAME_PROPERTIES( frame ), "motion_est.macroblock_width", c->macroblock_width );
-       mlt_properties_set_int( MLT_FRAME_PROPERTIES( frame ), "motion_est.macroblock_height", c->macroblock_height );
 
        return error;
 }
@@ -1053,24 +1081,23 @@ mlt_filter filter_motion_est_init( char *arg )
                this->process = filter_process;
 
                /* defaults that may be overridden */
-               context->macroblock_width = 16;
-               context->macroblock_height = 16;
+               context->mb_w = 16;
+               context->mb_h = 16;
                context->skip_prediction = 0;
                context->limit_x = 64;
                context->limit_y = 64;
-               context->search_method = DIAMOND_SEARCH;
+               context->search_method = DIAMOND_SEARCH;        // FIXME: not used
                context->check_chroma = 0;
                context->denoise = 1;
+               context->show_reconstruction = 0;
+               context->show_residual = 0;
+               context->toggle_when_paused = 0;
 
                /* reference functions that may have optimized versions */
                context->compare_reference = sad_reference;
-               //context->vert_deviation_reference = vertical_gradient_reference;
-               //context->horiz_deviation_reference = horizontal_gradient_reference;
 
                // The rest of the buffers will be initialized when the filter is first processed
                context->initialized = 0;
        }
        return this;
 }
-
-/** This source code will self destruct in 5...4...3... */