History

lavfi: add filter dnn_detect for object detection

Below are the example steps to do object detection: 1. download and install l_openvino_toolkit_p_2021.1.110.tgz from https://software.intel.com/content/www/us/en/develop/tools/openvino-toolkit/download.html or, we can get source code (tag 2021.1), build and install. 2. export LD_LIBRARY_PATH with openvino settings, for example: .../deployment_tools/inference_engine/lib/intel64/:.../deployment_tools/inference_engine/external/tbb/lib/ 3. rebuild ffmpeg from source code with configure option: --enable-libopenvino --extra-cflags='-I.../deployment_tools/inference_engine/include/' --extra-ldflags='-L.../deployment_tools/inference_engine/lib/intel64' 4. download model files and test image wget https://github.com/guoyejun/ffmpeg_dnn/raw/main/models/openvino/2021.1/face-detection-adas-0001.bin wget https://github.com/guoyejun/ffmpeg_dnn/raw/main/models/openvino/2021.1/face-detection-adas-0001.xml wget https://github.com/guoyejun/ffmpeg_dnn/raw/main/models/openvino/2021.1/face-detection-adas-0001.label wget https://github.com/guoyejun/ffmpeg_dnn/raw/main/images/cici.jpg 5. run ffmpeg with: ./ffmpeg -i cici.jpg -vf dnn_detect=dnn_backend=openvino:model=face-detection-adas-0001.xml:input=data:output=detection_out:confidence=0.6:labels=face-detection-adas-0001.label,showinfo -f null - We'll see the detect result as below: [Parsed_showinfo_1 @ 0x560c21ecbe40] side data - detection bounding boxes: [Parsed_showinfo_1 @ 0x560c21ecbe40] source: face-detection-adas-0001.xml [Parsed_showinfo_1 @ 0x560c21ecbe40] index: 0, region: (1005, 813) -> (1086, 905), label: face, confidence: 10000/10000. [Parsed_showinfo_1 @ 0x560c21ecbe40] index: 1, region: (888, 839) -> (967, 926), label: face, confidence: 6917/10000. There are two faces detected with confidence 100% and 69.17%.

Guo, Yejun committed 5 years ago

Commit aa9ffda1 parent e942b4b5 files · +464 −0Browse files

Unified · Split

5 files
  1. configure +1 −0
  2. doc/filters.texi +40 −0
  3. libavfilter/Makefile +1 −0
  4. libavfilter/allfilters.c +1 −0
  5. libavfilter/vf_dnn_detect.c +421 −0
configure+1−0
@@ -3555,6 +3555,7 @@ derain_filter_select="dnn" · 3554 unmodified lines
35553555 deshake_filter_select="pixelutils"
35563556 deshake_opencl_filter_deps="opencl"
35573557 dilation_opencl_filter_deps="opencl"
3558+dnn_detect_filter_select="dnn"
35583559 dnn_processing_filter_select="dnn"
35593560 drawtext_filter_deps="libfreetype"
35603561 drawtext_filter_suggest="libfontconfig libfribidi"
doc/filters.texi+40−0
@@ -10127,6 +10127,46 @@ ffmpeg -i INPUT -f lavfi -i nullsrc=hd720,geq='r=128+80*(sin(sqrt((X-W/2)*(X-W/2 · 10126 unmodified lines
1012710127 @end example
1012810128 @end itemize
1012910129
10130+@section dnn_detect
10131+
10132+Do object detection with deep neural networks.
10133+
10134+The filter accepts the following options:
10135+
10136+@table @option
10137+@item dnn_backend
10138+Specify which DNN backend to use for model loading and execution. This option accepts
10139+only openvino now, tensorflow backends will be added.
10140+
10141+@item model
10142+Set path to model file specifying network architecture and its parameters.
10143+Note that different backends use different file formats.
10144+
10145+@item input
10146+Set the input name of the dnn network.
10147+
10148+@item output
10149+Set the output name of the dnn network.
10150+
10151+@item confidence
10152+Set the confidence threshold (default: 0.5).
10153+
10154+@item labels
10155+Set path to label file specifying the mapping between label id and name.
10156+Each label name is written in one line, tailing spaces and empty lines are skipped.
10157+The first line is the name of label id 0 (usually it is 'background'),
10158+and the second line is the name of label id 1, etc.
10159+The label id is considered as name if the label file is not provided.
10160+
10161+@item backend_configs
10162+Set the configs to be passed into backend
10163+
10164+@item async
10165+use DNN async execution if set (default: set),
10166+roll back to sync execution if the backend does not support async.
10167+
10168+@end table
10169+
1013010170 @anchor{dnn_processing}
1013110171 @section dnn_processing
1013210172
libavfilter/Makefile+1−0
@@ -245,6 +245,7 @@ OBJS-$(CONFIG_DILATION_FILTER) += vf_neighbor.o · 244 unmodified lines
245245 OBJS-$(CONFIG_DILATION_OPENCL_FILTER) += vf_neighbor_opencl.o opencl.o \
246246 opencl/neighbor.o
247247 OBJS-$(CONFIG_DISPLACE_FILTER) += vf_displace.o framesync.o
248+OBJS-$(CONFIG_DNN_DETECT_FILTER) += vf_dnn_detect.o
248249 OBJS-$(CONFIG_DNN_PROCESSING_FILTER) += vf_dnn_processing.o
249250 OBJS-$(CONFIG_DOUBLEWEAVE_FILTER) += vf_weave.o
250251 OBJS-$(CONFIG_DRAWBOX_FILTER) += vf_drawbox.o
libavfilter/allfilters.c+1−0
@@ -230,6 +230,7 @@ extern AVFilter ff_vf_detelecine; · 229 unmodified lines
230230 extern AVFilter ff_vf_dilation;
231231 extern AVFilter ff_vf_dilation_opencl;
232232 extern AVFilter ff_vf_displace;
233+extern AVFilter ff_vf_dnn_detect;
233234 extern AVFilter ff_vf_dnn_processing;
234235 extern AVFilter ff_vf_doubleweave;
235236 extern AVFilter ff_vf_drawbox;
libavfilter/vf_dnn_detect.c+421−0
@@ -0,0 +1,421 @@
1+/*
2+ * This file is part of FFmpeg.
3+ *
4+ * FFmpeg is free software; you can redistribute it and/or
5+ * modify it under the terms of the GNU Lesser General Public
6+ * License as published by the Free Software Foundation; either
7+ * version 2.1 of the License, or (at your option) any later version.
8+ *
9+ * FFmpeg is distributed in the hope that it will be useful,
10+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
11+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12+ * Lesser General Public License for more details.
13+ *
14+ * You should have received a copy of the GNU Lesser General Public
15+ * License along with FFmpeg; if not, write to the Free Software
16+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
17+ */
18+
19+/**
20+ * @file
21+ * implementing an object detecting filter using deep learning networks.
22+ */
23+
24+#include "libavformat/avio.h"
25+#include "libavutil/opt.h"
26+#include "libavutil/pixdesc.h"
27+#include "libavutil/avassert.h"
28+#include "libavutil/imgutils.h"
29+#include "filters.h"
30+#include "dnn_filter_common.h"
31+#include "formats.h"
32+#include "internal.h"
33+#include "libavutil/time.h"
34+#include "libavutil/avstring.h"
35+#include "libavutil/detection_bbox.h"
36+
37+typedef struct DnnDetectContext {
38+ const AVClass *class;
39+ DnnContext dnnctx;
40+ float confidence;
41+ char *labels_filename;
42+ char **labels;
43+ int label_count;
44+} DnnDetectContext;
45+
46+#define OFFSET(x) offsetof(DnnDetectContext, dnnctx.x)
47+#define OFFSET2(x) offsetof(DnnDetectContext, x)
48+#define FLAGS AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM
49+static const AVOption dnn_detect_options[] = {
50+ { "dnn_backend", "DNN backend", OFFSET(backend_type), AV_OPT_TYPE_INT, { .i64 = 2 }, INT_MIN, INT_MAX, FLAGS, "backend" },
51+#if (CONFIG_LIBOPENVINO == 1)
52+ { "openvino", "openvino backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = 2 }, 0, 0, FLAGS, "backend" },
53+#endif
54+ DNN_COMMON_OPTIONS
55+ { "confidence", "threshold of confidence", OFFSET2(confidence), AV_OPT_TYPE_FLOAT, { .dbl = 0.5 }, 0, 1, FLAGS},
56+ { "labels", "path to labels file", OFFSET2(labels_filename), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, FLAGS },
57+ { NULL }
58+};
59+
60+AVFILTER_DEFINE_CLASS(dnn_detect);
61+
62+static int dnn_detect_post_proc(AVFrame *frame, DNNData *output, uint32_t nb, AVFilterContext *filter_ctx)
63+{
64+ DnnDetectContext *ctx = filter_ctx->priv;
65+ float conf_threshold = ctx->confidence;
66+ int proposal_count = output->height;
67+ int detect_size = output->width;
68+ float *detections = output->data;
69+ int nb_bboxes = 0;
70+ AVFrameSideData *sd;
71+ AVDetectionBBox *bbox;
72+ AVDetectionBBoxHeader *header;
73+
74+ sd = av_frame_get_side_data(frame, AV_FRAME_DATA_DETECTION_BBOXES);
75+ if (sd) {
76+ av_log(filter_ctx, AV_LOG_ERROR, "already have bounding boxes in side data.\n");
77+ return -1;
78+ }
79+
80+ for (int i = 0; i < proposal_count; ++i) {
81+ float conf = detections[i * detect_size + 2];
82+ if (conf < conf_threshold) {
83+ continue;
84+ }
85+ nb_bboxes++;
86+ }
87+
88+ if (nb_bboxes == 0) {
89+ av_log(filter_ctx, AV_LOG_VERBOSE, "nothing detected in this frame.\n");
90+ return 0;
91+ }
92+
93+ header = av_detection_bbox_create_side_data(frame, nb_bboxes);
94+ if (!header) {
95+ av_log(filter_ctx, AV_LOG_ERROR, "failed to create side data with %d bounding boxes\n", nb_bboxes);
96+ return -1;
97+ }
98+
99+ av_strlcpy(header->source, ctx->dnnctx.model_filename, sizeof(header->source));
100+
101+ for (int i = 0; i < proposal_count; ++i) {
102+ int av_unused image_id = (int)detections[i * detect_size + 0];
103+ int label_id = (int)detections[i * detect_size + 1];
104+ float conf = detections[i * detect_size + 2];
105+ float x0 = detections[i * detect_size + 3];
106+ float y0 = detections[i * detect_size + 4];
107+ float x1 = detections[i * detect_size + 5];
108+ float y1 = detections[i * detect_size + 6];
109+
110+ bbox = av_get_detection_bbox(header, i);
111+
112+ if (conf < conf_threshold) {
113+ continue;
114+ }
115+
116+ bbox->x = (int)(x0 * frame->width);
117+ bbox->w = (int)(x1 * frame->width) - bbox->x;
118+ bbox->y = (int)(y0 * frame->height);
119+ bbox->h = (int)(y1 * frame->height) - bbox->y;
120+
121+ bbox->detect_confidence = av_make_q((int)(conf * 10000), 10000);
122+ bbox->classify_count = 0;
123+
124+ if (ctx->labels && label_id < ctx->label_count) {
125+ av_strlcpy(bbox->detect_label, ctx->labels[label_id], sizeof(bbox->detect_label));
126+ } else {
127+ snprintf(bbox->detect_label, sizeof(bbox->detect_label), "%d", label_id);
128+ }
129+
130+ nb_bboxes--;
131+ if (nb_bboxes == 0) {
132+ break;
133+ }
134+ }
135+
136+ return 0;
137+}
138+
139+static void free_detect_labels(DnnDetectContext *ctx)
140+{
141+ for (int i = 0; i < ctx->label_count; i++) {
142+ av_freep(&ctx->labels[i]);
143+ }
144+ ctx->label_count = 0;
145+ av_freep(&ctx->labels);
146+}
147+
148+static int read_detect_label_file(AVFilterContext *context)
149+{
150+ int line_len;
151+ FILE *file;
152+ DnnDetectContext *ctx = context->priv;
153+
154+ file = av_fopen_utf8(ctx->labels_filename, "r");
155+ if (!file){
156+ av_log(context, AV_LOG_ERROR, "failed to open file %s\n", ctx->labels_filename);
157+ return AVERROR(EINVAL);
158+ }
159+
160+ while (!feof(file)) {
161+ char *label;
162+ char buf[256];
163+ if (!fgets(buf, 256, file)) {
164+ break;
165+ }
166+
167+ line_len = strlen(buf);
168+ while (line_len) {
169+ int i = line_len - 1;
170+ if (buf[i] == '\n' || buf[i] == '\r' || buf[i] == ' ') {
171+ buf[i] = '\0';
172+ line_len--;
173+ } else {
174+ break;
175+ }
176+ }
177+
178+ if (line_len == 0) // empty line
179+ continue;
180+
181+ if (line_len >= AV_DETECTION_BBOX_LABEL_NAME_MAX_SIZE) {
182+ av_log(context, AV_LOG_ERROR, "label %s too long\n", buf);
183+ fclose(file);
184+ return AVERROR(EINVAL);
185+ }
186+
187+ label = av_strdup(buf);
188+ if (!label) {
189+ av_log(context, AV_LOG_ERROR, "failed to allocate memory for label %s\n", buf);
190+ fclose(file);
191+ return AVERROR(ENOMEM);
192+ }
193+
194+ if (av_dynarray_add_nofree(&ctx->labels, &ctx->label_count, label) < 0) {
195+ av_log(context, AV_LOG_ERROR, "failed to do av_dynarray_add\n");
196+ fclose(file);
197+ av_freep(&label);
198+ return AVERROR(ENOMEM);
199+ }
200+ }
201+
202+ fclose(file);
203+ return 0;
204+}
205+
206+static av_cold int dnn_detect_init(AVFilterContext *context)
207+{
208+ DnnDetectContext *ctx = context->priv;
209+ int ret = ff_dnn_init(&ctx->dnnctx, DFT_ANALYTICS_DETECT, context);
210+ if (ret < 0)
211+ return ret;
212+ ff_dnn_set_detect_post_proc(&ctx->dnnctx, dnn_detect_post_proc);
213+
214+ if (ctx->labels_filename) {
215+ return read_detect_label_file(context);
216+ }
217+ return 0;
218+}
219+
220+static int dnn_detect_query_formats(AVFilterContext *context)
221+{
222+ static const enum AVPixelFormat pix_fmts[] = {
223+ AV_PIX_FMT_RGB24, AV_PIX_FMT_BGR24,
224+ AV_PIX_FMT_GRAY8, AV_PIX_FMT_GRAYF32,
225+ AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
226+ AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
227+ AV_PIX_FMT_NV12,
228+ AV_PIX_FMT_NONE
229+ };
230+ AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
231+ return ff_set_common_formats(context, fmts_list);
232+}
233+
234+static int dnn_detect_filter_frame(AVFilterLink *inlink, AVFrame *in)
235+{
236+ AVFilterContext *context = inlink->dst;
237+ AVFilterLink *outlink = context->outputs[0];
238+ DnnDetectContext *ctx = context->priv;
239+ DNNReturnType dnn_result;
240+
241+ dnn_result = ff_dnn_execute_model(&ctx->dnnctx, in, in);
242+ if (dnn_result != DNN_SUCCESS){
243+ av_log(ctx, AV_LOG_ERROR, "failed to execute model\n");
244+ av_frame_free(&in);
245+ return AVERROR(EIO);
246+ }
247+
248+ return ff_filter_frame(outlink, in);
249+}
250+
251+static int dnn_detect_activate_sync(AVFilterContext *filter_ctx)
252+{
253+ AVFilterLink *inlink = filter_ctx->inputs[0];
254+ AVFilterLink *outlink = filter_ctx->outputs[0];
255+ AVFrame *in = NULL;
256+ int64_t pts;
257+ int ret, status;
258+ int got_frame = 0;
259+
260+ FF_FILTER_FORWARD_STATUS_BACK(outlink, inlink);
261+
262+ do {
263+ // drain all input frames
264+ ret = ff_inlink_consume_frame(inlink, &in);
265+ if (ret < 0)
266+ return ret;
267+ if (ret > 0) {
268+ ret = dnn_detect_filter_frame(inlink, in);
269+ if (ret < 0)
270+ return ret;
271+ got_frame = 1;
272+ }
273+ } while (ret > 0);
274+
275+ // if frame got, schedule to next filter
276+ if (got_frame)
277+ return 0;
278+
279+ if (ff_inlink_acknowledge_status(inlink, &status, &pts)) {
280+ if (status == AVERROR_EOF) {
281+ ff_outlink_set_status(outlink, status, pts);
282+ return ret;
283+ }
284+ }
285+
286+ FF_FILTER_FORWARD_WANTED(outlink, inlink);
287+
288+ return FFERROR_NOT_READY;
289+}
290+
291+static int dnn_detect_flush_frame(AVFilterLink *outlink, int64_t pts, int64_t *out_pts)
292+{
293+ DnnDetectContext *ctx = outlink->src->priv;
294+ int ret;
295+ DNNAsyncStatusType async_state;
296+
297+ ret = ff_dnn_flush(&ctx->dnnctx);
298+ if (ret != DNN_SUCCESS) {
299+ return -1;
300+ }
301+
302+ do {
303+ AVFrame *in_frame = NULL;
304+ AVFrame *out_frame = NULL;
305+ async_state = ff_dnn_get_async_result(&ctx->dnnctx, &in_frame, &out_frame);
306+ if (out_frame) {
307+ av_assert0(in_frame == out_frame);
308+ ret = ff_filter_frame(outlink, out_frame);
309+ if (ret < 0)
310+ return ret;
311+ if (out_pts)
312+ *out_pts = out_frame->pts + pts;
313+ }
314+ av_usleep(5000);
315+ } while (async_state >= DAST_NOT_READY);
316+
317+ return 0;
318+}
319+
320+static int dnn_detect_activate_async(AVFilterContext *filter_ctx)
321+{
322+ AVFilterLink *inlink = filter_ctx->inputs[0];
323+ AVFilterLink *outlink = filter_ctx->outputs[0];
324+ DnnDetectContext *ctx = filter_ctx->priv;
325+ AVFrame *in = NULL;
326+ int64_t pts;
327+ int ret, status;
328+ int got_frame = 0;
329+ int async_state;
330+
331+ FF_FILTER_FORWARD_STATUS_BACK(outlink, inlink);
332+
333+ do {
334+ // drain all input frames
335+ ret = ff_inlink_consume_frame(inlink, &in);
336+ if (ret < 0)
337+ return ret;
338+ if (ret > 0) {
339+ if (ff_dnn_execute_model_async(&ctx->dnnctx, in, in) != DNN_SUCCESS) {
340+ return AVERROR(EIO);
341+ }
342+ }
343+ } while (ret > 0);
344+
345+ // drain all processed frames
346+ do {
347+ AVFrame *in_frame = NULL;
348+ AVFrame *out_frame = NULL;
349+ async_state = ff_dnn_get_async_result(&ctx->dnnctx, &in_frame, &out_frame);
350+ if (out_frame) {
351+ av_assert0(in_frame == out_frame);
352+ ret = ff_filter_frame(outlink, out_frame);
353+ if (ret < 0)
354+ return ret;
355+ got_frame = 1;
356+ }
357+ } while (async_state == DAST_SUCCESS);
358+
359+ // if frame got, schedule to next filter
360+ if (got_frame)
361+ return 0;
362+
363+ if (ff_inlink_acknowledge_status(inlink, &status, &pts)) {
364+ if (status == AVERROR_EOF) {
365+ int64_t out_pts = pts;
366+ ret = dnn_detect_flush_frame(outlink, pts, &out_pts);
367+ ff_outlink_set_status(outlink, status, out_pts);
368+ return ret;
369+ }
370+ }
371+
372+ FF_FILTER_FORWARD_WANTED(outlink, inlink);
373+
374+ return 0;
375+}
376+
377+static int dnn_detect_activate(AVFilterContext *filter_ctx)
378+{
379+ DnnDetectContext *ctx = filter_ctx->priv;
380+
381+ if (ctx->dnnctx.async)
382+ return dnn_detect_activate_async(filter_ctx);
383+ else
384+ return dnn_detect_activate_sync(filter_ctx);
385+}
386+
387+static av_cold void dnn_detect_uninit(AVFilterContext *context)
388+{
389+ DnnDetectContext *ctx = context->priv;
390+ ff_dnn_uninit(&ctx->dnnctx);
391+ free_detect_labels(ctx);
392+}
393+
394+static const AVFilterPad dnn_detect_inputs[] = {
395+ {
396+ .name = "default",
397+ .type = AVMEDIA_TYPE_VIDEO,
398+ },
399+ { NULL }
400+};
401+
402+static const AVFilterPad dnn_detect_outputs[] = {
403+ {
404+ .name = "default",
405+ .type = AVMEDIA_TYPE_VIDEO,
406+ },
407+ { NULL }
408+};
409+
410+AVFilter ff_vf_dnn_detect = {
411+ .name = "dnn_detect",
412+ .description = NULL_IF_CONFIG_SMALL("Apply DNN detect filter to the input."),
413+ .priv_size = sizeof(DnnDetectContext),
414+ .init = dnn_detect_init,
415+ .uninit = dnn_detect_uninit,
416+ .query_formats = dnn_detect_query_formats,
417+ .inputs = dnn_detect_inputs,
418+ .outputs = dnn_detect_outputs,
419+ .priv_class = &dnn_detect_class,
420+ .activate = dnn_detect_activate,
421+};

Keyboard shortcuts

General

Show keyboard shortcuts
?
Focus the current search
/

Issues and pull requests

Open filters
f

Issues, pull requests and Inbox

Select next item
j↓
Select previous item
k↑
Open selected item
Enter
Dismiss filters or clear selection
Esc

Repository code

Go to file
t