FFmpeg
vf_dnn_classify.c
Go to the documentation of this file.
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 classification filter using deep learning networks.
22  */
23 
24 #include "libavutil/file_open.h"
25 #include "libavutil/opt.h"
26 #include "filters.h"
27 #include "dnn_filter_common.h"
28 #include "internal.h"
29 #include "video.h"
30 #include "libavutil/time.h"
31 #include "libavutil/avstring.h"
33 
34 typedef struct DnnClassifyContext {
35  const AVClass *class;
37  float confidence;
39  char *target;
40  char **labels;
43 
44 #define OFFSET(x) offsetof(DnnClassifyContext, dnnctx.x)
45 #define OFFSET2(x) offsetof(DnnClassifyContext, x)
46 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM
47 static const AVOption dnn_classify_options[] = {
48  { "dnn_backend", "DNN backend", OFFSET(backend_type), AV_OPT_TYPE_INT, { .i64 = DNN_OV }, INT_MIN, INT_MAX, FLAGS, .unit = "backend" },
49 #if (CONFIG_LIBOPENVINO == 1)
50  { "openvino", "openvino backend flag", 0, AV_OPT_TYPE_CONST, { .i64 = DNN_OV }, 0, 0, FLAGS, .unit = "backend" },
51 #endif
53  { "confidence", "threshold of confidence", OFFSET2(confidence), AV_OPT_TYPE_FLOAT, { .dbl = 0.5 }, 0, 1, FLAGS},
54  { "labels", "path to labels file", OFFSET2(labels_filename), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, FLAGS },
55  { "target", "which one to be classified", OFFSET2(target), AV_OPT_TYPE_STRING, { .str = NULL }, 0, 0, FLAGS },
56  { NULL }
57 };
58 
59 AVFILTER_DEFINE_CLASS(dnn_classify);
60 
62 {
64  float conf_threshold = ctx->confidence;
66  AVDetectionBBox *bbox;
67  float *classifications;
68  uint32_t label_id;
69  float confidence;
70  AVFrameSideData *sd;
71  int output_size = output->dims[3] * output->dims[2] * output->dims[1];
72  if (output_size <= 0) {
73  return -1;
74  }
75 
77  if (!sd) {
78  av_log(filter_ctx, AV_LOG_ERROR, "Cannot get side data in dnn_classify_post_proc\n");
79  return -1;
80  }
82 
83  if (bbox_index == 0) {
84  av_strlcat(header->source, ", ", sizeof(header->source));
85  av_strlcat(header->source, ctx->dnnctx.model_filename, sizeof(header->source));
86  }
87 
88  classifications = output->data;
89  label_id = 0;
90  confidence= classifications[0];
91  for (int i = 1; i < output_size; i++) {
92  if (classifications[i] > confidence) {
93  label_id = i;
94  confidence= classifications[i];
95  }
96  }
97 
98  if (confidence < conf_threshold) {
99  return 0;
100  }
101 
102  bbox = av_get_detection_bbox(header, bbox_index);
103  bbox->classify_confidences[bbox->classify_count] = av_make_q((int)(confidence * 10000), 10000);
104 
105  if (ctx->labels && label_id < ctx->label_count) {
106  av_strlcpy(bbox->classify_labels[bbox->classify_count], ctx->labels[label_id], sizeof(bbox->classify_labels[bbox->classify_count]));
107  } else {
108  snprintf(bbox->classify_labels[bbox->classify_count], sizeof(bbox->classify_labels[bbox->classify_count]), "%d", label_id);
109  }
110 
111  bbox->classify_count++;
112 
113  return 0;
114 }
115 
117 {
118  for (int i = 0; i < ctx->label_count; i++) {
119  av_freep(&ctx->labels[i]);
120  }
121  ctx->label_count = 0;
122  av_freep(&ctx->labels);
123 }
124 
126 {
127  int line_len;
128  FILE *file;
129  DnnClassifyContext *ctx = context->priv;
130 
131  file = avpriv_fopen_utf8(ctx->labels_filename, "r");
132  if (!file){
133  av_log(context, AV_LOG_ERROR, "failed to open file %s\n", ctx->labels_filename);
134  return AVERROR(EINVAL);
135  }
136 
137  while (!feof(file)) {
138  char *label;
139  char buf[256];
140  if (!fgets(buf, 256, file)) {
141  break;
142  }
143 
144  line_len = strlen(buf);
145  while (line_len) {
146  int i = line_len - 1;
147  if (buf[i] == '\n' || buf[i] == '\r' || buf[i] == ' ') {
148  buf[i] = '\0';
149  line_len--;
150  } else {
151  break;
152  }
153  }
154 
155  if (line_len == 0) // empty line
156  continue;
157 
158  if (line_len >= AV_DETECTION_BBOX_LABEL_NAME_MAX_SIZE) {
159  av_log(context, AV_LOG_ERROR, "label %s too long\n", buf);
160  fclose(file);
161  return AVERROR(EINVAL);
162  }
163 
164  label = av_strdup(buf);
165  if (!label) {
166  av_log(context, AV_LOG_ERROR, "failed to allocate memory for label %s\n", buf);
167  fclose(file);
168  return AVERROR(ENOMEM);
169  }
170 
171  if (av_dynarray_add_nofree(&ctx->labels, &ctx->label_count, label) < 0) {
172  av_log(context, AV_LOG_ERROR, "failed to do av_dynarray_add\n");
173  fclose(file);
174  av_freep(&label);
175  return AVERROR(ENOMEM);
176  }
177  }
178 
179  fclose(file);
180  return 0;
181 }
182 
184 {
185  DnnClassifyContext *ctx = context->priv;
187  if (ret < 0)
188  return ret;
190 
191  if (ctx->labels_filename) {
193  }
194  return 0;
195 }
196 
197 static const enum AVPixelFormat pix_fmts[] = {
204 };
205 
206 static int dnn_classify_flush_frame(AVFilterLink *outlink, int64_t pts, int64_t *out_pts)
207 {
208  DnnClassifyContext *ctx = outlink->src->priv;
209  int ret;
210  DNNAsyncStatusType async_state;
211 
212  ret = ff_dnn_flush(&ctx->dnnctx);
213  if (ret != 0) {
214  return -1;
215  }
216 
217  do {
218  AVFrame *in_frame = NULL;
219  AVFrame *out_frame = NULL;
220  async_state = ff_dnn_get_result(&ctx->dnnctx, &in_frame, &out_frame);
221  if (async_state == DAST_SUCCESS) {
222  ret = ff_filter_frame(outlink, in_frame);
223  if (ret < 0)
224  return ret;
225  if (out_pts)
226  *out_pts = in_frame->pts + pts;
227  }
228  av_usleep(5000);
229  } while (async_state >= DAST_NOT_READY);
230 
231  return 0;
232 }
233 
235 {
236  AVFilterLink *inlink = filter_ctx->inputs[0];
237  AVFilterLink *outlink = filter_ctx->outputs[0];
239  AVFrame *in = NULL;
240  int64_t pts;
241  int ret, status;
242  int got_frame = 0;
243  int async_state;
244 
246 
247  do {
248  // drain all input frames
250  if (ret < 0)
251  return ret;
252  if (ret > 0) {
253  if (ff_dnn_execute_model_classification(&ctx->dnnctx, in, NULL, ctx->target) != 0) {
254  return AVERROR(EIO);
255  }
256  }
257  } while (ret > 0);
258 
259  // drain all processed frames
260  do {
261  AVFrame *in_frame = NULL;
262  AVFrame *out_frame = NULL;
263  async_state = ff_dnn_get_result(&ctx->dnnctx, &in_frame, &out_frame);
264  if (async_state == DAST_SUCCESS) {
265  ret = ff_filter_frame(outlink, in_frame);
266  if (ret < 0)
267  return ret;
268  got_frame = 1;
269  }
270  } while (async_state == DAST_SUCCESS);
271 
272  // if frame got, schedule to next filter
273  if (got_frame)
274  return 0;
275 
277  if (status == AVERROR_EOF) {
278  int64_t out_pts = pts;
279  ret = dnn_classify_flush_frame(outlink, pts, &out_pts);
280  ff_outlink_set_status(outlink, status, out_pts);
281  return ret;
282  }
283  }
284 
286 
287  return 0;
288 }
289 
291 {
292  DnnClassifyContext *ctx = context->priv;
293  ff_dnn_uninit(&ctx->dnnctx);
295 }
296 
298  .name = "dnn_classify",
299  .description = NULL_IF_CONFIG_SMALL("Apply DNN classify filter to the input."),
300  .priv_size = sizeof(DnnClassifyContext),
306  .priv_class = &dnn_classify_class,
307  .activate = dnn_classify_activate,
308 };
dnn_classify_flush_frame
static int dnn_classify_flush_frame(AVFilterLink *outlink, int64_t pts, int64_t *out_pts)
Definition: vf_dnn_classify.c:206
free_classify_labels
static void free_classify_labels(DnnClassifyContext *ctx)
Definition: vf_dnn_classify.c:116
AVPixelFormat
AVPixelFormat
Pixel format.
Definition: pixfmt.h:71
AVERROR
Filter the word “frame” indicates either a video frame or a group of audio as stored in an AVFrame structure Format for each input and each output the list of supported formats For video that means pixel format For audio that means channel sample they are references to shared objects When the negotiation mechanism computes the intersection of the formats supported at each end of a all references to both lists are replaced with a reference to the intersection And when a single format is eventually chosen for a link amongst the remaining all references to the list are updated That means that if a filter requires that its input and output have the same format amongst a supported all it has to do is use a reference to the same list of formats query_formats can leave some formats unset and return AVERROR(EAGAIN) to cause the negotiation mechanism toagain later. That can be used by filters with complex requirements to use the format negotiated on one link to set the formats supported on another. Frame references ownership and permissions
opt.h
av_frame_get_side_data
AVFrameSideData * av_frame_get_side_data(const AVFrame *frame, enum AVFrameSideDataType type)
Definition: frame.c:858
ff_filter_frame
int ff_filter_frame(AVFilterLink *link, AVFrame *frame)
Send a frame of data to the next filter.
Definition: avfilter.c:1018
AVERROR_EOF
#define AVERROR_EOF
End of file.
Definition: error.h:57
FILTER_PIXFMTS_ARRAY
#define FILTER_PIXFMTS_ARRAY(array)
Definition: internal.h:162
output
filter_frame For filters that do not use the this method is called when a frame is pushed to the filter s input It can be called at any time except in a reentrant way If the input frame is enough to produce output
Definition: filter_design.txt:225
inlink
The exact code depends on how similar the blocks are and how related they are to the and needs to apply these operations to the correct inlink or outlink if there are several Macros are available to factor that when no extra processing is inlink
Definition: filter_design.txt:212
AVFrame
This structure describes decoded (raw) audio or video data.
Definition: frame.h:375
AVFrame::pts
int64_t pts
Presentation timestamp in time_base units (time when frame should be shown to user).
Definition: frame.h:487
AVOption
AVOption.
Definition: opt.h:346
DnnClassifyContext::confidence
float confidence
Definition: vf_dnn_classify.c:37
AV_PIX_FMT_BGR24
@ AV_PIX_FMT_BGR24
packed RGB 8:8:8, 24bpp, BGRBGR...
Definition: pixfmt.h:76
AVFILTER_DEFINE_CLASS
AVFILTER_DEFINE_CLASS(dnn_classify)
AVFilter::name
const char * name
Filter name.
Definition: avfilter.h:170
dnn_filter_common.h
video.h
FF_FILTER_FORWARD_STATUS_BACK
#define FF_FILTER_FORWARD_STATUS_BACK(outlink, inlink)
Forward the status on an output link to an input link.
Definition: filters.h:199
ff_inlink_consume_frame
int ff_inlink_consume_frame(AVFilterLink *link, AVFrame **rframe)
Take a frame from the link's FIFO and update the link's stats.
Definition: avfilter.c:1445
AV_DETECTION_BBOX_LABEL_NAME_MAX_SIZE
#define AV_DETECTION_BBOX_LABEL_NAME_MAX_SIZE
Definition: detection_bbox.h:36
AVFilterContext::priv
void * priv
private data for use by the filter
Definition: avfilter.h:422
dnn_classify_post_proc
static int dnn_classify_post_proc(AVFrame *frame, DNNData *output, uint32_t bbox_index, AVFilterContext *filter_ctx)
Definition: vf_dnn_classify.c:61
DnnContext
Definition: dnn_filter_common.h:29
filter_ctx
static FilteringContext * filter_ctx
Definition: transcode.c:51
pts
static int64_t pts
Definition: transcode_aac.c:643
av_get_detection_bbox
static av_always_inline AVDetectionBBox * av_get_detection_bbox(const AVDetectionBBoxHeader *header, unsigned int idx)
Definition: detection_bbox.h:84
AV_LOG_ERROR
#define AV_LOG_ERROR
Something went wrong and cannot losslessly be recovered.
Definition: log.h:180
dnn_classify_options
static const AVOption dnn_classify_options[]
Definition: vf_dnn_classify.c:47
av_cold
#define av_cold
Definition: attributes.h:90
ff_video_default_filterpad
const AVFilterPad ff_video_default_filterpad[1]
An AVFilterPad array whose only entry has name "default" and is of type AVMEDIA_TYPE_VIDEO.
Definition: video.c:37
ff_outlink_set_status
static void ff_outlink_set_status(AVFilterLink *link, int status, int64_t pts)
Set the status field of a link from the source filter.
Definition: filters.h:189
AVDetectionBBox::classify_confidences
AVRational classify_confidences[AV_NUM_DETECTION_BBOX_CLASSIFY]
Definition: detection_bbox.h:53
DNNData
Definition: dnn_interface.h:65
filters.h
ff_dnn_get_result
DNNAsyncStatusType ff_dnn_get_result(DnnContext *ctx, AVFrame **in_frame, AVFrame **out_frame)
Definition: dnn_filter_common.c:161
DnnClassifyContext
Definition: vf_dnn_classify.c:34
ctx
AVFormatContext * ctx
Definition: movenc.c:48
AV_PIX_FMT_YUV420P
@ AV_PIX_FMT_YUV420P
planar YUV 4:2:0, 12bpp, (1 Cr & Cb sample per 2x2 Y samples)
Definition: pixfmt.h:73
av_usleep
int av_usleep(unsigned usec)
Sleep for a period of time.
Definition: time.c:84
OFFSET2
#define OFFSET2(x)
Definition: vf_dnn_classify.c:45
AV_PIX_FMT_GRAYF32
#define AV_PIX_FMT_GRAYF32
Definition: pixfmt.h:511
FILTER_INPUTS
#define FILTER_INPUTS(array)
Definition: internal.h:182
file_open.h
frame
static AVFrame * frame
Definition: demux_decode.c:54
DNN_OV
@ DNN_OV
Definition: dnn_interface.h:35
if
if(ret)
Definition: filter_design.txt:179
context
it s the only field you need to keep assuming you have a context There is some magic you don t need to care about around this just let it vf default minimum maximum flags name is the option keep it simple and lowercase description are in without and describe what they for example set the foo of the bar offset is the offset of the field in your context
Definition: writing_filters.txt:91
AVClass
Describe the class of an AVClass context structure.
Definition: log.h:66
NULL
#define NULL
Definition: coverity.c:32
AVDetectionBBox::classify_labels
char classify_labels[AV_NUM_DETECTION_BBOX_CLASSIFY][AV_DETECTION_BBOX_LABEL_NAME_MAX_SIZE]
Definition: detection_bbox.h:52
AVDetectionBBoxHeader
Definition: detection_bbox.h:56
ff_dnn_execute_model_classification
int ff_dnn_execute_model_classification(DnnContext *ctx, AVFrame *in_frame, AVFrame *out_frame, const char *target)
Definition: dnn_filter_common.c:146
dnn_classify_init
static av_cold int dnn_classify_init(AVFilterContext *context)
Definition: vf_dnn_classify.c:183
time.h
DnnClassifyContext::target
char * target
Definition: vf_dnn_classify.c:39
AV_PIX_FMT_GRAY8
@ AV_PIX_FMT_GRAY8
Y , 8bpp.
Definition: pixfmt.h:81
ff_dnn_flush
int ff_dnn_flush(DnnContext *ctx)
Definition: dnn_filter_common.c:166
ff_inlink_acknowledge_status
int ff_inlink_acknowledge_status(AVFilterLink *link, int *rstatus, int64_t *rpts)
Test and acknowledge the change of status on the link.
Definition: avfilter.c:1392
DNN_COMMON_OPTIONS
#define DNN_COMMON_OPTIONS
Definition: dnn_filter_common.h:43
init
int(* init)(AVBSFContext *ctx)
Definition: dts2pts.c:365
AV_PIX_FMT_RGB24
@ AV_PIX_FMT_RGB24
packed RGB 8:8:8, 24bpp, RGBRGB...
Definition: pixfmt.h:75
NULL_IF_CONFIG_SMALL
#define NULL_IF_CONFIG_SMALL(x)
Return NULL if CONFIG_SMALL is true, otherwise the argument without modification.
Definition: internal.h:106
DAST_SUCCESS
@ DAST_SUCCESS
Definition: dnn_interface.h:49
av_make_q
static AVRational av_make_q(int num, int den)
Create an AVRational.
Definition: rational.h:71
AVFrameSideData::data
uint8_t * data
Definition: frame.h:252
DnnClassifyContext::label_count
int label_count
Definition: vf_dnn_classify.c:41
header
static const uint8_t header[24]
Definition: sdr2.c:68
AVDetectionBBox::classify_count
uint32_t classify_count
Definition: detection_bbox.h:51
FF_FILTER_FORWARD_WANTED
FF_FILTER_FORWARD_WANTED(outlink, inlink)
DnnClassifyContext::labels
char ** labels
Definition: vf_dnn_classify.c:40
internal.h
AV_OPT_TYPE_FLOAT
@ AV_OPT_TYPE_FLOAT
Definition: opt.h:238
ff_vf_dnn_classify
const AVFilter ff_vf_dnn_classify
Definition: vf_dnn_classify.c:297
uninit
static void uninit(AVBSFContext *ctx)
Definition: pcm_rechunk.c:68
i
#define i(width, name, range_min, range_max)
Definition: cbs_h2645.c:255
dnn_classify_uninit
static av_cold void dnn_classify_uninit(AVFilterContext *context)
Definition: vf_dnn_classify.c:290
avpriv_fopen_utf8
FILE * avpriv_fopen_utf8(const char *path, const char *mode)
Open a file using a UTF-8 filename.
Definition: file_open.c:159
AVFilter
Filter definition.
Definition: avfilter.h:166
ret
ret
Definition: filter_design.txt:187
AV_PIX_FMT_NV12
@ AV_PIX_FMT_NV12
planar YUV 4:2:0, 12bpp, 1 plane for Y and 1 plane for the UV components, which are interleaved (firs...
Definition: pixfmt.h:96
av_strlcat
size_t av_strlcat(char *dst, const char *src, size_t size)
Append the string src to the string dst, but to a total length of no more than size - 1 bytes,...
Definition: avstring.c:95
OFFSET
#define OFFSET(x)
Definition: vf_dnn_classify.c:44
av_dynarray_add_nofree
int av_dynarray_add_nofree(void *tab_ptr, int *nb_ptr, void *elem)
Add an element to a dynamic array.
Definition: mem.c:313
DFT_ANALYTICS_CLASSIFY
@ DFT_ANALYTICS_CLASSIFY
Definition: dnn_interface.h:56
status
ov_status_e status
Definition: dnn_backend_openvino.c:120
AV_PIX_FMT_NONE
@ AV_PIX_FMT_NONE
Definition: pixfmt.h:72
AV_OPT_TYPE_INT
@ AV_OPT_TYPE_INT
Definition: opt.h:235
AV_PIX_FMT_YUV444P
@ AV_PIX_FMT_YUV444P
planar YUV 4:4:4, 24bpp, (1 Cr & Cb sample per 1x1 Y samples)
Definition: pixfmt.h:78
AVFilterContext
An instance of a filter.
Definition: avfilter.h:407
pix_fmts
static enum AVPixelFormat pix_fmts[]
Definition: vf_dnn_classify.c:197
av_strdup
char * av_strdup(const char *s)
Duplicate a string.
Definition: mem.c:270
AV_PIX_FMT_YUV422P
@ AV_PIX_FMT_YUV422P
planar YUV 4:2:2, 16bpp, (1 Cr & Cb sample per 2x1 Y samples)
Definition: pixfmt.h:77
AVFrameSideData
Structure to hold side data for an AVFrame.
Definition: frame.h:250
FILTER_OUTPUTS
#define FILTER_OUTPUTS(array)
Definition: internal.h:183
FLAGS
#define FLAGS
Definition: vf_dnn_classify.c:46
ff_dnn_init
int ff_dnn_init(DnnContext *ctx, DNNFunctionType func_type, AVFilterContext *filter_ctx)
Definition: dnn_filter_common.c:54
av_freep
#define av_freep(p)
Definition: tableprint_vlc.h:34
AV_PIX_FMT_YUV411P
@ AV_PIX_FMT_YUV411P
planar YUV 4:1:1, 12bpp, (1 Cr & Cb sample per 4x1 Y samples)
Definition: pixfmt.h:80
dnn_classify_activate
static int dnn_classify_activate(AVFilterContext *filter_ctx)
Definition: vf_dnn_classify.c:234
AV_PIX_FMT_YUV410P
@ AV_PIX_FMT_YUV410P
planar YUV 4:1:0, 9bpp, (1 Cr & Cb sample per 4x4 Y samples)
Definition: pixfmt.h:79
av_strlcpy
size_t av_strlcpy(char *dst, const char *src, size_t size)
Copy the string src to dst, but no more than size - 1 bytes, and null-terminate dst.
Definition: avstring.c:85
av_log
#define av_log(a,...)
Definition: tableprint_vlc.h:27
ff_dnn_uninit
void ff_dnn_uninit(DnnContext *ctx)
Definition: dnn_filter_common.c:171
AVDetectionBBox
Definition: detection_bbox.h:26
read_classify_label_file
static int read_classify_label_file(AVFilterContext *context)
Definition: vf_dnn_classify.c:125
avstring.h
AV_OPT_TYPE_STRING
@ AV_OPT_TYPE_STRING
Definition: opt.h:239
DAST_NOT_READY
@ DAST_NOT_READY
Definition: dnn_interface.h:48
DNNAsyncStatusType
DNNAsyncStatusType
Definition: dnn_interface.h:45
AV_OPT_TYPE_CONST
@ AV_OPT_TYPE_CONST
Definition: opt.h:244
snprintf
#define snprintf
Definition: snprintf.h:34
detection_bbox.h
ff_dnn_set_classify_post_proc
int ff_dnn_set_classify_post_proc(DnnContext *ctx, ClassifyPostProc post_proc)
Definition: dnn_filter_common.c:115
AV_FRAME_DATA_DETECTION_BBOXES
@ AV_FRAME_DATA_DETECTION_BBOXES
Bounding boxes for object detection and classification, as described by AVDetectionBBoxHeader.
Definition: frame.h:194
DnnClassifyContext::labels_filename
char * labels_filename
Definition: vf_dnn_classify.c:38
DnnClassifyContext::dnnctx
DnnContext dnnctx
Definition: vf_dnn_classify.c:36