Line |
Branch |
Exec |
Source |
1 |
|
|
/* |
2 |
|
|
* Copyright (C) 2010 Georg Martius <georg.martius@web.de> |
3 |
|
|
* Copyright (C) 2010 Daniel G. Taylor <dan@programmer-art.org> |
4 |
|
|
* |
5 |
|
|
* This file is part of FFmpeg. |
6 |
|
|
* |
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. |
11 |
|
|
* |
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. |
16 |
|
|
* |
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 |
20 |
|
|
*/ |
21 |
|
|
|
22 |
|
|
/** |
23 |
|
|
* @file |
24 |
|
|
* fast deshake / depan video filter |
25 |
|
|
* |
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. |
29 |
|
|
* |
30 |
|
|
* Algorithm: |
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 |
38 |
|
|
* |
39 |
|
|
* TODO: |
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? |
43 |
|
|
* |
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 |
47 |
|
|
* |
48 |
|
|
* Special thanks to Jason Kotenko for his help with the algorithm and my |
49 |
|
|
* inability to see simple errors in C code. |
50 |
|
|
*/ |
51 |
|
|
|
52 |
|
|
#include "avfilter.h" |
53 |
|
|
#include "filters.h" |
54 |
|
|
#include "transform.h" |
55 |
|
|
#include "video.h" |
56 |
|
|
#include "libavutil/common.h" |
57 |
|
|
#include "libavutil/file_open.h" |
58 |
|
|
#include "libavutil/mem.h" |
59 |
|
|
#include "libavutil/opt.h" |
60 |
|
|
#include "libavutil/pixdesc.h" |
61 |
|
|
#include "libavutil/pixelutils.h" |
62 |
|
|
#include "libavutil/qsort.h" |
63 |
|
|
|
64 |
|
|
|
65 |
|
|
enum SearchMethod { |
66 |
|
|
EXHAUSTIVE, ///< Search all possible positions |
67 |
|
|
SMART_EXHAUSTIVE, ///< Search most possible positions (faster) |
68 |
|
|
SEARCH_COUNT |
69 |
|
|
}; |
70 |
|
|
|
71 |
|
|
typedef struct IntMotionVector { |
72 |
|
|
int x; ///< Horizontal shift |
73 |
|
|
int y; ///< Vertical shift |
74 |
|
|
} IntMotionVector; |
75 |
|
|
|
76 |
|
|
typedef struct MotionVector { |
77 |
|
|
double x; ///< Horizontal shift |
78 |
|
|
double y; ///< Vertical shift |
79 |
|
|
} MotionVector; |
80 |
|
|
|
81 |
|
|
typedef struct Transform { |
82 |
|
|
MotionVector vec; ///< Motion vector |
83 |
|
|
double angle; ///< Angle of rotation |
84 |
|
|
double zoom; ///< Zoom percentage |
85 |
|
|
} Transform; |
86 |
|
|
|
87 |
|
|
#define MAX_R 64 |
88 |
|
|
|
89 |
|
|
typedef struct DeshakeContext { |
90 |
|
|
const AVClass *class; |
91 |
|
|
int counts[2*MAX_R+1][2*MAX_R+1]; ///< Scratch buffer for motion search |
92 |
|
|
double *angles; ///< Scratch buffer for block angles |
93 |
|
|
unsigned angles_size; |
94 |
|
|
AVFrame *ref; ///< Previous frame |
95 |
|
|
int rx; ///< Maximum horizontal shift |
96 |
|
|
int ry; ///< Maximum vertical shift |
97 |
|
|
int edge; ///< Edge fill method |
98 |
|
|
int blocksize; ///< Size of blocks to compare |
99 |
|
|
int contrast; ///< Contrast threshold |
100 |
|
|
int search; ///< Motion search method |
101 |
|
|
av_pixelutils_sad_fn sad; ///< Sum of the absolute difference function |
102 |
|
|
Transform last; ///< Transform from last frame |
103 |
|
|
int refcount; ///< Number of reference frames (defines averaging window) |
104 |
|
|
FILE *fp; |
105 |
|
|
Transform avg; |
106 |
|
|
int cw; ///< Crop motion search to this box |
107 |
|
|
int ch; |
108 |
|
|
int cx; |
109 |
|
|
int cy; |
110 |
|
|
char *filename; ///< Motion search detailed log filename |
111 |
|
|
int opencl; |
112 |
|
|
int (* transform)(AVFilterContext *ctx, int width, int height, int cw, int ch, |
113 |
|
|
const float *matrix_y, const float *matrix_uv, enum InterpolateMethod interpolate, |
114 |
|
|
enum FillMethod fill, AVFrame *in, AVFrame *out); |
115 |
|
|
} DeshakeContext; |
116 |
|
|
|
117 |
|
|
#define OFFSET(x) offsetof(DeshakeContext, x) |
118 |
|
|
#define FLAGS AV_OPT_FLAG_VIDEO_PARAM|AV_OPT_FLAG_FILTERING_PARAM |
119 |
|
|
|
120 |
|
|
static const AVOption deshake_options[] = { |
121 |
|
|
{ "x", "set x for the rectangular search area", OFFSET(cx), AV_OPT_TYPE_INT, {.i64=-1}, -1, INT_MAX, .flags = FLAGS }, |
122 |
|
|
{ "y", "set y for the rectangular search area", OFFSET(cy), AV_OPT_TYPE_INT, {.i64=-1}, -1, INT_MAX, .flags = FLAGS }, |
123 |
|
|
{ "w", "set width for the rectangular search area", OFFSET(cw), AV_OPT_TYPE_INT, {.i64=-1}, -1, INT_MAX, .flags = FLAGS }, |
124 |
|
|
{ "h", "set height for the rectangular search area", OFFSET(ch), AV_OPT_TYPE_INT, {.i64=-1}, -1, INT_MAX, .flags = FLAGS }, |
125 |
|
|
{ "rx", "set x for the rectangular search area", OFFSET(rx), AV_OPT_TYPE_INT, {.i64=16}, 0, MAX_R, .flags = FLAGS }, |
126 |
|
|
{ "ry", "set y for the rectangular search area", OFFSET(ry), AV_OPT_TYPE_INT, {.i64=16}, 0, MAX_R, .flags = FLAGS }, |
127 |
|
|
{ "edge", "set edge mode", OFFSET(edge), AV_OPT_TYPE_INT, {.i64=FILL_MIRROR}, FILL_BLANK, FILL_COUNT-1, FLAGS, .unit = "edge"}, |
128 |
|
|
{ "blank", "fill zeroes at blank locations", 0, AV_OPT_TYPE_CONST, {.i64=FILL_BLANK}, INT_MIN, INT_MAX, FLAGS, .unit = "edge" }, |
129 |
|
|
{ "original", "original image at blank locations", 0, AV_OPT_TYPE_CONST, {.i64=FILL_ORIGINAL}, INT_MIN, INT_MAX, FLAGS, .unit = "edge" }, |
130 |
|
|
{ "clamp", "extruded edge value at blank locations", 0, AV_OPT_TYPE_CONST, {.i64=FILL_CLAMP}, INT_MIN, INT_MAX, FLAGS, .unit = "edge" }, |
131 |
|
|
{ "mirror", "mirrored edge at blank locations", 0, AV_OPT_TYPE_CONST, {.i64=FILL_MIRROR}, INT_MIN, INT_MAX, FLAGS, .unit = "edge" }, |
132 |
|
|
{ "blocksize", "set motion search blocksize", OFFSET(blocksize), AV_OPT_TYPE_INT, {.i64=8}, 4, 128, .flags = FLAGS }, |
133 |
|
|
{ "contrast", "set contrast threshold for blocks", OFFSET(contrast), AV_OPT_TYPE_INT, {.i64=125}, 1, 255, .flags = FLAGS }, |
134 |
|
|
{ "search", "set search strategy", OFFSET(search), AV_OPT_TYPE_INT, {.i64=EXHAUSTIVE}, EXHAUSTIVE, SEARCH_COUNT-1, FLAGS, .unit = "smode" }, |
135 |
|
|
{ "exhaustive", "exhaustive search", 0, AV_OPT_TYPE_CONST, {.i64=EXHAUSTIVE}, INT_MIN, INT_MAX, FLAGS, .unit = "smode" }, |
136 |
|
|
{ "less", "less exhaustive search", 0, AV_OPT_TYPE_CONST, {.i64=SMART_EXHAUSTIVE}, INT_MIN, INT_MAX, FLAGS, .unit = "smode" }, |
137 |
|
|
{ "filename", "set motion search detailed log file name", OFFSET(filename), AV_OPT_TYPE_STRING, {.str=NULL}, .flags = FLAGS }, |
138 |
|
|
{ "opencl", "ignored", OFFSET(opencl), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, .flags = FLAGS }, |
139 |
|
|
{ NULL } |
140 |
|
|
}; |
141 |
|
|
|
142 |
|
|
AVFILTER_DEFINE_CLASS(deshake); |
143 |
|
|
|
144 |
|
✗ |
static int cmp(const void *a, const void *b) |
145 |
|
|
{ |
146 |
|
✗ |
return FFDIFFSIGN(*(const double *)a, *(const double *)b); |
147 |
|
|
} |
148 |
|
|
|
149 |
|
|
/** |
150 |
|
|
* Cleaned mean (cuts off 20% of values to remove outliers and then averages) |
151 |
|
|
*/ |
152 |
|
✗ |
static double clean_mean(double *values, int count) |
153 |
|
|
{ |
154 |
|
✗ |
double mean = 0; |
155 |
|
✗ |
int cut = count / 5; |
156 |
|
|
int x; |
157 |
|
|
|
158 |
|
✗ |
AV_QSORT(values, count, double, cmp); |
159 |
|
|
|
160 |
|
✗ |
for (x = cut; x < count - cut; x++) { |
161 |
|
✗ |
mean += values[x]; |
162 |
|
|
} |
163 |
|
|
|
164 |
|
✗ |
return mean / (count - cut * 2); |
165 |
|
|
} |
166 |
|
|
|
167 |
|
|
/** |
168 |
|
|
* Find the most likely shift in motion between two frames for a given |
169 |
|
|
* macroblock. Test each block against several shifts given by the rx |
170 |
|
|
* and ry attributes. Searches using a simple matrix of those shifts and |
171 |
|
|
* chooses the most likely shift by the smallest difference in blocks. |
172 |
|
|
*/ |
173 |
|
✗ |
static void find_block_motion(DeshakeContext *deshake, uint8_t *src1, |
174 |
|
|
uint8_t *src2, int cx, int cy, int stride, |
175 |
|
|
IntMotionVector *mv) |
176 |
|
|
{ |
177 |
|
|
int x, y; |
178 |
|
|
int diff; |
179 |
|
✗ |
int smallest = INT_MAX; |
180 |
|
|
int tmp, tmp2; |
181 |
|
|
|
182 |
|
|
#define CMP(i, j) deshake->sad(src1 + cy * stride + cx, stride,\ |
183 |
|
|
src2 + (j) * stride + (i), stride) |
184 |
|
|
|
185 |
|
✗ |
if (deshake->search == EXHAUSTIVE) { |
186 |
|
|
// Compare every possible position - this is sloooow! |
187 |
|
✗ |
for (y = -deshake->ry; y <= deshake->ry; y++) { |
188 |
|
✗ |
for (x = -deshake->rx; x <= deshake->rx; x++) { |
189 |
|
✗ |
diff = CMP(cx - x, cy - y); |
190 |
|
✗ |
if (diff < smallest) { |
191 |
|
✗ |
smallest = diff; |
192 |
|
✗ |
mv->x = x; |
193 |
|
✗ |
mv->y = y; |
194 |
|
|
} |
195 |
|
|
} |
196 |
|
|
} |
197 |
|
✗ |
} else if (deshake->search == SMART_EXHAUSTIVE) { |
198 |
|
|
// Compare every other possible position and find the best match |
199 |
|
✗ |
for (y = -deshake->ry + 1; y < deshake->ry; y += 2) { |
200 |
|
✗ |
for (x = -deshake->rx + 1; x < deshake->rx; x += 2) { |
201 |
|
✗ |
diff = CMP(cx - x, cy - y); |
202 |
|
✗ |
if (diff < smallest) { |
203 |
|
✗ |
smallest = diff; |
204 |
|
✗ |
mv->x = x; |
205 |
|
✗ |
mv->y = y; |
206 |
|
|
} |
207 |
|
|
} |
208 |
|
|
} |
209 |
|
|
|
210 |
|
|
// Hone in on the specific best match around the match we found above |
211 |
|
✗ |
tmp = mv->x; |
212 |
|
✗ |
tmp2 = mv->y; |
213 |
|
|
|
214 |
|
✗ |
for (y = tmp2 - 1; y <= tmp2 + 1; y++) { |
215 |
|
✗ |
for (x = tmp - 1; x <= tmp + 1; x++) { |
216 |
|
✗ |
if (x == tmp && y == tmp2) |
217 |
|
✗ |
continue; |
218 |
|
|
|
219 |
|
✗ |
diff = CMP(cx - x, cy - y); |
220 |
|
✗ |
if (diff < smallest) { |
221 |
|
✗ |
smallest = diff; |
222 |
|
✗ |
mv->x = x; |
223 |
|
✗ |
mv->y = y; |
224 |
|
|
} |
225 |
|
|
} |
226 |
|
|
} |
227 |
|
|
} |
228 |
|
|
|
229 |
|
✗ |
if (smallest > 512) { |
230 |
|
✗ |
mv->x = -1; |
231 |
|
✗ |
mv->y = -1; |
232 |
|
|
} |
233 |
|
|
//av_log(NULL, AV_LOG_ERROR, "%d\n", smallest); |
234 |
|
|
//av_log(NULL, AV_LOG_ERROR, "Final: (%d, %d) = %d x %d\n", cx, cy, mv->x, mv->y); |
235 |
|
✗ |
} |
236 |
|
|
|
237 |
|
|
/** |
238 |
|
|
* Find the contrast of a given block. When searching for global motion we |
239 |
|
|
* really only care about the high contrast blocks, so using this method we |
240 |
|
|
* can actually skip blocks we don't care much about. |
241 |
|
|
*/ |
242 |
|
✗ |
static int block_contrast(uint8_t *src, int x, int y, int stride, int blocksize) |
243 |
|
|
{ |
244 |
|
✗ |
int highest = 0; |
245 |
|
✗ |
int lowest = 255; |
246 |
|
|
int i, j, pos; |
247 |
|
|
|
248 |
|
✗ |
for (i = 0; i <= blocksize * 2; i++) { |
249 |
|
|
// We use a width of 16 here to match the sad function |
250 |
|
✗ |
for (j = 0; j <= 15; j++) { |
251 |
|
✗ |
pos = (y + i) * stride + (x + j); |
252 |
|
✗ |
if (src[pos] < lowest) |
253 |
|
✗ |
lowest = src[pos]; |
254 |
|
✗ |
else if (src[pos] > highest) { |
255 |
|
✗ |
highest = src[pos]; |
256 |
|
|
} |
257 |
|
|
} |
258 |
|
|
} |
259 |
|
|
|
260 |
|
✗ |
return highest - lowest; |
261 |
|
|
} |
262 |
|
|
|
263 |
|
|
/** |
264 |
|
|
* Find the rotation for a given block. |
265 |
|
|
*/ |
266 |
|
✗ |
static double block_angle(int x, int y, int cx, int cy, IntMotionVector *shift) |
267 |
|
|
{ |
268 |
|
|
double a1, a2, diff; |
269 |
|
|
|
270 |
|
✗ |
a1 = atan2(y - cy, x - cx); |
271 |
|
✗ |
a2 = atan2(y - cy + shift->y, x - cx + shift->x); |
272 |
|
|
|
273 |
|
✗ |
diff = a2 - a1; |
274 |
|
|
|
275 |
|
✗ |
return (diff > M_PI) ? diff - 2 * M_PI : |
276 |
|
✗ |
(diff < -M_PI) ? diff + 2 * M_PI : |
277 |
|
|
diff; |
278 |
|
|
} |
279 |
|
|
|
280 |
|
|
/** |
281 |
|
|
* Find the estimated global motion for a scene given the most likely shift |
282 |
|
|
* for each block in the frame. The global motion is estimated to be the |
283 |
|
|
* same as the motion from most blocks in the frame, so if most blocks |
284 |
|
|
* move one pixel to the right and two pixels down, this would yield a |
285 |
|
|
* motion vector (1, -2). |
286 |
|
|
*/ |
287 |
|
✗ |
static void find_motion(DeshakeContext *deshake, uint8_t *src1, uint8_t *src2, |
288 |
|
|
int width, int height, int stride, Transform *t) |
289 |
|
|
{ |
290 |
|
|
int x, y; |
291 |
|
✗ |
IntMotionVector mv = {0, 0}; |
292 |
|
✗ |
int count_max_value = 0; |
293 |
|
|
int contrast; |
294 |
|
|
|
295 |
|
|
int pos; |
296 |
|
✗ |
int center_x = 0, center_y = 0; |
297 |
|
|
double p_x, p_y; |
298 |
|
|
|
299 |
|
✗ |
av_fast_malloc(&deshake->angles, &deshake->angles_size, width * height / (16 * deshake->blocksize) * sizeof(*deshake->angles)); |
300 |
|
|
|
301 |
|
|
// Reset counts to zero |
302 |
|
✗ |
for (x = 0; x < deshake->rx * 2 + 1; x++) { |
303 |
|
✗ |
for (y = 0; y < deshake->ry * 2 + 1; y++) { |
304 |
|
✗ |
deshake->counts[x][y] = 0; |
305 |
|
|
} |
306 |
|
|
} |
307 |
|
|
|
308 |
|
✗ |
pos = 0; |
309 |
|
|
// Find motion for every block and store the motion vector in the counts |
310 |
|
✗ |
for (y = deshake->ry; y < height - deshake->ry - (deshake->blocksize * 2); y += deshake->blocksize * 2) { |
311 |
|
|
// We use a width of 16 here to match the sad function |
312 |
|
✗ |
for (x = deshake->rx; x < width - deshake->rx - 16; x += 16) { |
313 |
|
|
// If the contrast is too low, just skip this block as it probably |
314 |
|
|
// won't be very useful to us. |
315 |
|
✗ |
contrast = block_contrast(src2, x, y, stride, deshake->blocksize); |
316 |
|
✗ |
if (contrast > deshake->contrast) { |
317 |
|
|
//av_log(NULL, AV_LOG_ERROR, "%d\n", contrast); |
318 |
|
✗ |
find_block_motion(deshake, src1, src2, x, y, stride, &mv); |
319 |
|
✗ |
if (mv.x != -1 && mv.y != -1) { |
320 |
|
✗ |
deshake->counts[mv.x + deshake->rx][mv.y + deshake->ry] += 1; |
321 |
|
✗ |
if (x > deshake->rx && y > deshake->ry) |
322 |
|
✗ |
deshake->angles[pos++] = block_angle(x, y, 0, 0, &mv); |
323 |
|
|
|
324 |
|
✗ |
center_x += mv.x; |
325 |
|
✗ |
center_y += mv.y; |
326 |
|
|
} |
327 |
|
|
} |
328 |
|
|
} |
329 |
|
|
} |
330 |
|
|
|
331 |
|
✗ |
if (pos) { |
332 |
|
✗ |
center_x /= pos; |
333 |
|
✗ |
center_y /= pos; |
334 |
|
✗ |
t->angle = clean_mean(deshake->angles, pos); |
335 |
|
✗ |
if (t->angle < 0.001) |
336 |
|
✗ |
t->angle = 0; |
337 |
|
|
} else { |
338 |
|
✗ |
t->angle = 0; |
339 |
|
|
} |
340 |
|
|
|
341 |
|
|
// Find the most common motion vector in the frame and use it as the gmv |
342 |
|
✗ |
for (y = deshake->ry * 2; y >= 0; y--) { |
343 |
|
✗ |
for (x = 0; x < deshake->rx * 2 + 1; x++) { |
344 |
|
|
//av_log(NULL, AV_LOG_ERROR, "%5d ", deshake->counts[x][y]); |
345 |
|
✗ |
if (deshake->counts[x][y] > count_max_value) { |
346 |
|
✗ |
t->vec.x = x - deshake->rx; |
347 |
|
✗ |
t->vec.y = y - deshake->ry; |
348 |
|
✗ |
count_max_value = deshake->counts[x][y]; |
349 |
|
|
} |
350 |
|
|
} |
351 |
|
|
//av_log(NULL, AV_LOG_ERROR, "\n"); |
352 |
|
|
} |
353 |
|
|
|
354 |
|
✗ |
p_x = (center_x - width / 2.0); |
355 |
|
✗ |
p_y = (center_y - height / 2.0); |
356 |
|
✗ |
t->vec.x += (cos(t->angle)-1)*p_x - sin(t->angle)*p_y; |
357 |
|
✗ |
t->vec.y += sin(t->angle)*p_x + (cos(t->angle)-1)*p_y; |
358 |
|
|
|
359 |
|
|
// Clamp max shift & rotation? |
360 |
|
✗ |
t->vec.x = av_clipf(t->vec.x, -deshake->rx * 2, deshake->rx * 2); |
361 |
|
✗ |
t->vec.y = av_clipf(t->vec.y, -deshake->ry * 2, deshake->ry * 2); |
362 |
|
✗ |
t->angle = av_clipf(t->angle, -0.1, 0.1); |
363 |
|
|
|
364 |
|
|
//av_log(NULL, AV_LOG_ERROR, "%d x %d\n", avg->x, avg->y); |
365 |
|
✗ |
} |
366 |
|
|
|
367 |
|
✗ |
static int deshake_transform_c(AVFilterContext *ctx, |
368 |
|
|
int width, int height, int cw, int ch, |
369 |
|
|
const float *matrix_y, const float *matrix_uv, |
370 |
|
|
enum InterpolateMethod interpolate, |
371 |
|
|
enum FillMethod fill, AVFrame *in, AVFrame *out) |
372 |
|
|
{ |
373 |
|
✗ |
int i = 0, ret = 0; |
374 |
|
|
const float *matrixs[3]; |
375 |
|
|
int plane_w[3], plane_h[3]; |
376 |
|
✗ |
matrixs[0] = matrix_y; |
377 |
|
✗ |
matrixs[1] = matrixs[2] = matrix_uv; |
378 |
|
✗ |
plane_w[0] = width; |
379 |
|
✗ |
plane_w[1] = plane_w[2] = cw; |
380 |
|
✗ |
plane_h[0] = height; |
381 |
|
✗ |
plane_h[1] = plane_h[2] = ch; |
382 |
|
|
|
383 |
|
✗ |
for (i = 0; i < 3; i++) { |
384 |
|
|
// Transform the luma and chroma planes |
385 |
|
✗ |
ret = ff_affine_transform(in->data[i], out->data[i], in->linesize[i], |
386 |
|
|
out->linesize[i], plane_w[i], plane_h[i], |
387 |
|
|
matrixs[i], interpolate, fill); |
388 |
|
✗ |
if (ret < 0) |
389 |
|
✗ |
return ret; |
390 |
|
|
} |
391 |
|
✗ |
return ret; |
392 |
|
|
} |
393 |
|
|
|
394 |
|
✗ |
static av_cold int init(AVFilterContext *ctx) |
395 |
|
|
{ |
396 |
|
✗ |
DeshakeContext *deshake = ctx->priv; |
397 |
|
|
|
398 |
|
✗ |
deshake->refcount = 20; // XXX: add to options? |
399 |
|
✗ |
deshake->blocksize /= 2; |
400 |
|
✗ |
deshake->blocksize = av_clip(deshake->blocksize, 4, 128); |
401 |
|
|
|
402 |
|
✗ |
if (deshake->rx % 16) { |
403 |
|
✗ |
av_log(ctx, AV_LOG_ERROR, "rx must be a multiple of 16\n"); |
404 |
|
✗ |
return AVERROR_PATCHWELCOME; |
405 |
|
|
} |
406 |
|
|
|
407 |
|
✗ |
if (deshake->filename) |
408 |
|
✗ |
deshake->fp = avpriv_fopen_utf8(deshake->filename, "w"); |
409 |
|
✗ |
if (deshake->fp) |
410 |
|
✗ |
fwrite("Ori x, Avg x, Fin x, Ori y, Avg y, Fin y, Ori angle, Avg angle, Fin angle, Ori zoom, Avg zoom, Fin zoom\n", 1, 104, deshake->fp); |
411 |
|
|
|
412 |
|
|
// Quadword align left edge of box for MMX code, adjust width if necessary |
413 |
|
|
// to keep right margin |
414 |
|
✗ |
if (deshake->cx > 0) { |
415 |
|
✗ |
deshake->cw += deshake->cx - (deshake->cx & ~15); |
416 |
|
✗ |
deshake->cx &= ~15; |
417 |
|
|
} |
418 |
|
✗ |
deshake->transform = deshake_transform_c; |
419 |
|
|
|
420 |
|
✗ |
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", |
421 |
|
|
deshake->cx, deshake->cy, deshake->cw, deshake->ch, |
422 |
|
✗ |
deshake->rx, deshake->ry, deshake->edge, deshake->blocksize * 2, deshake->contrast, deshake->search); |
423 |
|
|
|
424 |
|
✗ |
return 0; |
425 |
|
|
} |
426 |
|
|
|
427 |
|
|
static const enum AVPixelFormat pix_fmts[] = { |
428 |
|
|
AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P, AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV410P, |
429 |
|
|
AV_PIX_FMT_YUV411P, AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUVJ420P, AV_PIX_FMT_YUVJ422P, |
430 |
|
|
AV_PIX_FMT_YUVJ444P, AV_PIX_FMT_YUVJ440P, AV_PIX_FMT_NONE |
431 |
|
|
}; |
432 |
|
|
|
433 |
|
✗ |
static int config_props(AVFilterLink *link) |
434 |
|
|
{ |
435 |
|
✗ |
DeshakeContext *deshake = link->dst->priv; |
436 |
|
|
|
437 |
|
✗ |
deshake->ref = NULL; |
438 |
|
✗ |
deshake->last.vec.x = 0; |
439 |
|
✗ |
deshake->last.vec.y = 0; |
440 |
|
✗ |
deshake->last.angle = 0; |
441 |
|
✗ |
deshake->last.zoom = 0; |
442 |
|
|
|
443 |
|
✗ |
return 0; |
444 |
|
|
} |
445 |
|
|
|
446 |
|
✗ |
static av_cold void uninit(AVFilterContext *ctx) |
447 |
|
|
{ |
448 |
|
✗ |
DeshakeContext *deshake = ctx->priv; |
449 |
|
✗ |
av_frame_free(&deshake->ref); |
450 |
|
✗ |
av_freep(&deshake->angles); |
451 |
|
✗ |
deshake->angles_size = 0; |
452 |
|
✗ |
if (deshake->fp) |
453 |
|
✗ |
fclose(deshake->fp); |
454 |
|
✗ |
} |
455 |
|
|
|
456 |
|
✗ |
static int filter_frame(AVFilterLink *link, AVFrame *in) |
457 |
|
|
{ |
458 |
|
✗ |
DeshakeContext *deshake = link->dst->priv; |
459 |
|
✗ |
AVFilterLink *outlink = link->dst->outputs[0]; |
460 |
|
|
AVFrame *out; |
461 |
|
✗ |
Transform t = {{0},0}, orig = {{0},0}; |
462 |
|
|
float matrix_y[9], matrix_uv[9]; |
463 |
|
✗ |
float alpha = 2.0 / deshake->refcount; |
464 |
|
|
char tmp[256]; |
465 |
|
✗ |
int ret = 0; |
466 |
|
✗ |
const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(link->format); |
467 |
|
✗ |
const int chroma_width = AV_CEIL_RSHIFT(link->w, desc->log2_chroma_w); |
468 |
|
✗ |
const int chroma_height = AV_CEIL_RSHIFT(link->h, desc->log2_chroma_h); |
469 |
|
|
int aligned; |
470 |
|
|
float transform_zoom; |
471 |
|
|
|
472 |
|
✗ |
out = ff_get_video_buffer(outlink, outlink->w, outlink->h); |
473 |
|
✗ |
if (!out) { |
474 |
|
✗ |
av_frame_free(&in); |
475 |
|
✗ |
return AVERROR(ENOMEM); |
476 |
|
|
} |
477 |
|
✗ |
av_frame_copy_props(out, in); |
478 |
|
|
|
479 |
|
✗ |
aligned = !((intptr_t)in->data[0] & 15 | in->linesize[0] & 15); |
480 |
|
✗ |
deshake->sad = av_pixelutils_get_sad_fn(4, 4, aligned, deshake); // 16x16, 2nd source unaligned |
481 |
|
✗ |
if (!deshake->sad) { |
482 |
|
✗ |
ret = AVERROR(EINVAL); |
483 |
|
✗ |
goto fail; |
484 |
|
|
} |
485 |
|
|
|
486 |
|
✗ |
if (deshake->cx < 0 || deshake->cy < 0 || deshake->cw < 0 || deshake->ch < 0) { |
487 |
|
|
// Find the most likely global motion for the current frame |
488 |
|
✗ |
find_motion(deshake, (deshake->ref == NULL) ? in->data[0] : deshake->ref->data[0], in->data[0], link->w, link->h, in->linesize[0], &t); |
489 |
|
|
} else { |
490 |
|
✗ |
uint8_t *src1 = (deshake->ref == NULL) ? in->data[0] : deshake->ref->data[0]; |
491 |
|
✗ |
uint8_t *src2 = in->data[0]; |
492 |
|
|
|
493 |
|
✗ |
deshake->cx = FFMIN(deshake->cx, link->w); |
494 |
|
✗ |
deshake->cy = FFMIN(deshake->cy, link->h); |
495 |
|
|
|
496 |
|
✗ |
if ((unsigned)deshake->cx + (unsigned)deshake->cw > link->w) deshake->cw = link->w - deshake->cx; |
497 |
|
✗ |
if ((unsigned)deshake->cy + (unsigned)deshake->ch > link->h) deshake->ch = link->h - deshake->cy; |
498 |
|
|
|
499 |
|
|
// Quadword align right margin |
500 |
|
✗ |
deshake->cw &= ~15; |
501 |
|
|
|
502 |
|
✗ |
src1 += deshake->cy * in->linesize[0] + deshake->cx; |
503 |
|
✗ |
src2 += deshake->cy * in->linesize[0] + deshake->cx; |
504 |
|
|
|
505 |
|
✗ |
find_motion(deshake, src1, src2, deshake->cw, deshake->ch, in->linesize[0], &t); |
506 |
|
|
} |
507 |
|
|
|
508 |
|
|
|
509 |
|
|
// Copy transform so we can output it later to compare to the smoothed value |
510 |
|
✗ |
orig.vec.x = t.vec.x; |
511 |
|
✗ |
orig.vec.y = t.vec.y; |
512 |
|
✗ |
orig.angle = t.angle; |
513 |
|
✗ |
orig.zoom = t.zoom; |
514 |
|
|
|
515 |
|
|
// Generate a one-sided moving exponential average |
516 |
|
✗ |
deshake->avg.vec.x = alpha * t.vec.x + (1.0 - alpha) * deshake->avg.vec.x; |
517 |
|
✗ |
deshake->avg.vec.y = alpha * t.vec.y + (1.0 - alpha) * deshake->avg.vec.y; |
518 |
|
✗ |
deshake->avg.angle = alpha * t.angle + (1.0 - alpha) * deshake->avg.angle; |
519 |
|
✗ |
deshake->avg.zoom = alpha * t.zoom + (1.0 - alpha) * deshake->avg.zoom; |
520 |
|
|
|
521 |
|
|
// Remove the average from the current motion to detect the motion that |
522 |
|
|
// is not on purpose, just as jitter from bumping the camera |
523 |
|
✗ |
t.vec.x -= deshake->avg.vec.x; |
524 |
|
✗ |
t.vec.y -= deshake->avg.vec.y; |
525 |
|
✗ |
t.angle -= deshake->avg.angle; |
526 |
|
✗ |
t.zoom -= deshake->avg.zoom; |
527 |
|
|
|
528 |
|
|
// Invert the motion to undo it |
529 |
|
✗ |
t.vec.x *= -1; |
530 |
|
✗ |
t.vec.y *= -1; |
531 |
|
✗ |
t.angle *= -1; |
532 |
|
|
|
533 |
|
|
// Write statistics to file |
534 |
|
✗ |
if (deshake->fp) { |
535 |
|
✗ |
snprintf(tmp, 256, "%f, %f, %f, %f, %f, %f, %f, %f, %f, %f, %f, %f\n", orig.vec.x, deshake->avg.vec.x, t.vec.x, orig.vec.y, deshake->avg.vec.y, t.vec.y, orig.angle, deshake->avg.angle, t.angle, orig.zoom, deshake->avg.zoom, t.zoom); |
536 |
|
✗ |
fwrite(tmp, 1, strlen(tmp), deshake->fp); |
537 |
|
|
} |
538 |
|
|
|
539 |
|
|
// Turn relative current frame motion into absolute by adding it to the |
540 |
|
|
// last absolute motion |
541 |
|
✗ |
t.vec.x += deshake->last.vec.x; |
542 |
|
✗ |
t.vec.y += deshake->last.vec.y; |
543 |
|
✗ |
t.angle += deshake->last.angle; |
544 |
|
✗ |
t.zoom += deshake->last.zoom; |
545 |
|
|
|
546 |
|
|
// Shrink motion by 10% to keep things centered in the camera frame |
547 |
|
✗ |
t.vec.x *= 0.9; |
548 |
|
✗ |
t.vec.y *= 0.9; |
549 |
|
✗ |
t.angle *= 0.9; |
550 |
|
|
|
551 |
|
|
// Store the last absolute motion information |
552 |
|
✗ |
deshake->last.vec.x = t.vec.x; |
553 |
|
✗ |
deshake->last.vec.y = t.vec.y; |
554 |
|
✗ |
deshake->last.angle = t.angle; |
555 |
|
✗ |
deshake->last.zoom = t.zoom; |
556 |
|
|
|
557 |
|
✗ |
transform_zoom = 1.0 + t.zoom / 100.0; |
558 |
|
|
|
559 |
|
|
// Generate a luma transformation matrix |
560 |
|
✗ |
ff_get_matrix(t.vec.x, t.vec.y, t.angle, transform_zoom, transform_zoom, matrix_y); |
561 |
|
|
// Generate a chroma transformation matrix |
562 |
|
✗ |
ff_get_matrix(t.vec.x / (link->w / chroma_width), t.vec.y / (link->h / chroma_height), t.angle, transform_zoom, transform_zoom, matrix_uv); |
563 |
|
|
// Transform the luma and chroma planes |
564 |
|
✗ |
ret = deshake->transform(link->dst, link->w, link->h, chroma_width, chroma_height, |
565 |
|
✗ |
matrix_y, matrix_uv, INTERPOLATE_BILINEAR, deshake->edge, in, out); |
566 |
|
|
|
567 |
|
|
// Cleanup the old reference frame |
568 |
|
✗ |
av_frame_free(&deshake->ref); |
569 |
|
|
|
570 |
|
✗ |
if (ret < 0) |
571 |
|
✗ |
goto fail; |
572 |
|
|
|
573 |
|
|
// Store the current frame as the reference frame for calculating the |
574 |
|
|
// motion of the next frame |
575 |
|
✗ |
deshake->ref = in; |
576 |
|
|
|
577 |
|
✗ |
return ff_filter_frame(outlink, out); |
578 |
|
✗ |
fail: |
579 |
|
✗ |
av_frame_free(&out); |
580 |
|
✗ |
return ret; |
581 |
|
|
} |
582 |
|
|
|
583 |
|
|
static const AVFilterPad deshake_inputs[] = { |
584 |
|
|
{ |
585 |
|
|
.name = "default", |
586 |
|
|
.type = AVMEDIA_TYPE_VIDEO, |
587 |
|
|
.filter_frame = filter_frame, |
588 |
|
|
.config_props = config_props, |
589 |
|
|
}, |
590 |
|
|
}; |
591 |
|
|
|
592 |
|
|
const AVFilter ff_vf_deshake = { |
593 |
|
|
.name = "deshake", |
594 |
|
|
.description = NULL_IF_CONFIG_SMALL("Stabilize shaky video."), |
595 |
|
|
.priv_size = sizeof(DeshakeContext), |
596 |
|
|
.init = init, |
597 |
|
|
.uninit = uninit, |
598 |
|
|
FILTER_INPUTS(deshake_inputs), |
599 |
|
|
FILTER_OUTPUTS(ff_video_default_filterpad), |
600 |
|
|
FILTER_PIXFMTS_ARRAY(pix_fmts), |
601 |
|
|
.priv_class = &deshake_class, |
602 |
|
|
}; |
603 |
|
|
|