[FFmpeg-devel] [PATCH 10/10] libavfilter: add filter dnn_detect for object detection

Guo, Yejun yejun.guo at intel.com
Wed Feb 10 06:59:13 EET 2021


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 (2021.1 version), 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/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:conf=0.6,showinfo -f null -

We'll see the detect result as below:
[Parsed_showinfo_1 @ 0x55db3ffb60c0]   side data - Dnn bounding boxes:
[Parsed_showinfo_1 @ 0x55db3ffb60c0] index: 0, region: (330/672, 203/384) -> (356/672, 226/384), label: 1, confidence: 10000/10000.
[Parsed_showinfo_1 @ 0x55db3ffb60c0] index: 1, region: (291/672, 209/384) -> (317/672, 231/384), label: 1, confidence: 6917/10000.

There are two faces detected with confidence 100% and 69.17%, and
the input image size of the model is 672x384. The two bounding boxes
in this image are (330, 203)->(356, 226) and (291, 209)->(317, 231).

Since the orignal input image size is 2048x1536, so the two bounding
boxese in the original image are
(330/672*2048=1006, 203/384*1536=812) -> (1085, 904) and
(887, 836) -> (966, 924), and we can check them manually.

Signed-off-by: Guo, Yejun <yejun.guo at intel.com>

Next, we'll add tensorflow backend and update filter vf_drawbox etc
to visualize the detect result.
---
 configure                              |   1 +
 doc/filters.texi                       |  33 +++
 libavfilter/Makefile                   |   1 +
 libavfilter/allfilters.c               |   1 +
 libavfilter/dnn/dnn_backend_openvino.c |  12 +
 libavfilter/dnn_filter_common.c        |   7 +
 libavfilter/dnn_filter_common.h        |   1 +
 libavfilter/dnn_interface.h            |   6 +-
 libavfilter/vf_dnn_detect.c            | 356 +++++++++++++++++++++++++
 9 files changed, 416 insertions(+), 2 deletions(-)
 create mode 100644 libavfilter/vf_dnn_detect.c

diff --git a/configure b/configure
index a76c2ec4ae..2d2668571d 100755
--- a/configure
+++ b/configure
@@ -3548,6 +3548,7 @@ derain_filter_select="dnn"
 deshake_filter_select="pixelutils"
 deshake_opencl_filter_deps="opencl"
 dilation_opencl_filter_deps="opencl"
+dnn_detect_filter_select="dnn"
 dnn_processing_filter_select="dnn"
 drawtext_filter_deps="libfreetype"
 drawtext_filter_suggest="libfontconfig libfribidi"
diff --git a/doc/filters.texi b/doc/filters.texi
index 079bba9a1e..340402e650 100644
--- a/doc/filters.texi
+++ b/doc/filters.texi
@@ -10073,6 +10073,39 @@ ffmpeg -i INPUT -f lavfi -i nullsrc=hd720,geq='r=128+80*(sin(sqrt((X-W/2)*(X-W/2
 @end example
 @end itemize
 
+ at section dnn_detect
+
+Do object detection with deep neural networks.
+
+The filter accepts the following options:
+
+ at table @option
+ at item dnn_backend
+Specify which DNN backend to use for model loading and execution. This option accepts
+only openvino now, tensorflow backends will be added.
+
+ at item model
+Set path to model file specifying network architecture and its parameters.
+Note that different backends use different file formats.
+
+ at item input
+Set the input name of the dnn network.
+
+ at item output
+Set the output name of the dnn network.
+
+ at item conf
+Set the confidence threshold (default: 0.5).
+
+ at item backend_configs
+Set the configs to be passed into backend
+
+ at item async
+use DNN async execution if set (default: set),
+roll back to sync execution if the backend does not support async.
+
+ at end table
+
 @anchor{dnn_processing}
 @section dnn_processing
 
diff --git a/libavfilter/Makefile b/libavfilter/Makefile
index b43933be64..6c39e7111b 100644
--- a/libavfilter/Makefile
+++ b/libavfilter/Makefile
@@ -244,6 +244,7 @@ OBJS-$(CONFIG_DILATION_FILTER)               += vf_neighbor.o
 OBJS-$(CONFIG_DILATION_OPENCL_FILTER)        += vf_neighbor_opencl.o opencl.o \
                                                 opencl/neighbor.o
 OBJS-$(CONFIG_DISPLACE_FILTER)               += vf_displace.o framesync.o
+OBJS-$(CONFIG_DNN_DETECT_FILTER)             += vf_dnn_detect.o
 OBJS-$(CONFIG_DNN_PROCESSING_FILTER)         += vf_dnn_processing.o
 OBJS-$(CONFIG_DOUBLEWEAVE_FILTER)            += vf_weave.o
 OBJS-$(CONFIG_DRAWBOX_FILTER)                += vf_drawbox.o
diff --git a/libavfilter/allfilters.c b/libavfilter/allfilters.c
index 73d859ce5e..37bb276685 100644
--- a/libavfilter/allfilters.c
+++ b/libavfilter/allfilters.c
@@ -229,6 +229,7 @@ extern AVFilter ff_vf_detelecine;
 extern AVFilter ff_vf_dilation;
 extern AVFilter ff_vf_dilation_opencl;
 extern AVFilter ff_vf_displace;
+extern AVFilter ff_vf_dnn_detect;
 extern AVFilter ff_vf_dnn_processing;
 extern AVFilter ff_vf_doubleweave;
 extern AVFilter ff_vf_drawbox;
diff --git a/libavfilter/dnn/dnn_backend_openvino.c b/libavfilter/dnn/dnn_backend_openvino.c
index 5be053b7f8..928d84b744 100644
--- a/libavfilter/dnn/dnn_backend_openvino.c
+++ b/libavfilter/dnn/dnn_backend_openvino.c
@@ -621,6 +621,12 @@ DNNReturnType ff_dnn_execute_model_ov(const DNNModel *model, const char *input_n
         return DNN_ERROR;
     }
 
+    if (model->func_type != DFT_PROCESS_FRAME) {
+        if (!out_frame) {
+            out_frame = in_frame;
+        }
+    }
+
     if (nb_output != 1) {
         // currently, the filter does not need multiple outputs,
         // so we just pending the support until we really need it.
@@ -674,6 +680,12 @@ DNNReturnType ff_dnn_execute_model_async_ov(const DNNModel *model, const char *i
         return DNN_ERROR;
     }
 
+    if (model->func_type != DFT_PROCESS_FRAME) {
+        if (!out_frame) {
+            out_frame = in_frame;
+        }
+    }
+
     task = av_malloc(sizeof(*task));
     if (!task) {
         av_log(ctx, AV_LOG_ERROR, "unable to alloc memory for task item.\n");
diff --git a/libavfilter/dnn_filter_common.c b/libavfilter/dnn_filter_common.c
index 413adba406..92b696e710 100644
--- a/libavfilter/dnn_filter_common.c
+++ b/libavfilter/dnn_filter_common.c
@@ -64,6 +64,13 @@ int ff_dnn_init(DnnContext *ctx, DNNFunctionType func_type, AVFilterContext *fil
     return 0;
 }
 
+int ff_dnn_set_proc(DnnContext *ctx, PRE_POST_PROC pre_proc, PRE_POST_PROC post_proc)
+{
+    ctx->model->pre_proc = pre_proc;
+    ctx->model->post_proc = post_proc;
+    return 0;
+}
+
 DNNReturnType ff_dnn_get_input(DnnContext *ctx, DNNData *input)
 {
     return ctx->model->get_input(ctx->model->model, input, ctx->model_inputname);
diff --git a/libavfilter/dnn_filter_common.h b/libavfilter/dnn_filter_common.h
index 79c4d3efe3..0e88b88bdd 100644
--- a/libavfilter/dnn_filter_common.h
+++ b/libavfilter/dnn_filter_common.h
@@ -48,6 +48,7 @@ typedef struct DnnContext {
 
 
 int ff_dnn_init(DnnContext *ctx, DNNFunctionType func_type, AVFilterContext *filter_ctx);
+int ff_dnn_set_proc(DnnContext *ctx, PRE_POST_PROC pre_proc, PRE_POST_PROC post_proc);
 DNNReturnType ff_dnn_get_input(DnnContext *ctx, DNNData *input);
 DNNReturnType ff_dnn_get_output(DnnContext *ctx, int input_width, int input_height, int *output_width, int *output_height);
 DNNReturnType ff_dnn_execute_model(DnnContext *ctx, AVFrame *in_frame, AVFrame *out_frame);
diff --git a/libavfilter/dnn_interface.h b/libavfilter/dnn_interface.h
index d3a0c58a61..90a08129f4 100644
--- a/libavfilter/dnn_interface.h
+++ b/libavfilter/dnn_interface.h
@@ -63,6 +63,8 @@ typedef struct DNNData{
     DNNColorOrder order;
 } DNNData;
 
+typedef int (*PRE_POST_PROC)(AVFrame *frame, DNNData *model, AVFilterContext *filter_ctx);
+
 typedef struct DNNModel{
     // Stores model that can be different for different backends.
     void *model;
@@ -80,10 +82,10 @@ typedef struct DNNModel{
                                 const char *output_name, int *output_width, int *output_height);
     // set the pre process to transfer data from AVFrame to DNNData
     // the default implementation within DNN is used if it is not provided by the filter
-    int (*pre_proc)(AVFrame *frame_in, DNNData *model_input, AVFilterContext *filter_ctx);
+    PRE_POST_PROC pre_proc;
     // set the post process to transfer data from DNNData to AVFrame
     // the default implementation within DNN is used if it is not provided by the filter
-    int (*post_proc)(AVFrame *frame_out, DNNData *model_output, AVFilterContext *filter_ctx);
+    PRE_POST_PROC post_proc;
 } DNNModel;
 
 // Stores pointers to functions for loading, executing, freeing DNN models for one of the backends.
diff --git a/libavfilter/vf_dnn_detect.c b/libavfilter/vf_dnn_detect.c
new file mode 100644
index 0000000000..bac5035ae8
--- /dev/null
+++ b/libavfilter/vf_dnn_detect.c
@@ -0,0 +1,356 @@
+/*
+ * Copyright (c) 2021
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+/**
+ * @file
+ * implementing an object detecting filter using deep learning networks.
+ */
+
+#include "libavformat/avio.h"
+#include "libavutil/opt.h"
+#include "libavutil/pixdesc.h"
+#include "libavutil/avassert.h"
+#include "libavutil/imgutils.h"
+#include "filters.h"
+#include "dnn_filter_common.h"
+#include "formats.h"
+#include "internal.h"
+#include "libavutil/time.h"
+#include "libavutil/dnn_bbox.h"
+
+typedef struct DnnDetectContext {
+    const AVClass *class;
+    DnnContext dnnctx;
+    float conf;
+    int model_input_width;
+    int model_input_height;
+} DnnDetectContext;
+
+#define OFFSET(x) offsetof(DnnDetectContext, dnnctx.x)
+#define OFFSET2(x) offsetof(DnnDetectContext, x)
+#define FLAGS AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_VIDEO_PARAM
+static const AVOption dnn_detect_options[] = {
+    { "dnn_backend", "DNN backend",                OFFSET(backend_type),     AV_OPT_TYPE_INT,       { .i64 = 2 },    INT_MIN, INT_MAX, FLAGS, "backend" },
+#if (CONFIG_LIBOPENVINO == 1)
+    { "openvino",    "openvino backend flag",      0,                        AV_OPT_TYPE_CONST,     { .i64 = 2 },    0, 0, FLAGS, "backend" },
+#endif
+    DNN_COMMON_OPTIONS
+    { "conf",       "threshold of confidence",    OFFSET2(conf),             AV_OPT_TYPE_FLOAT,     { .dbl = 0.5 },  0, 1, FLAGS},
+    { NULL }
+};
+
+AVFILTER_DEFINE_CLASS(dnn_detect);
+
+static int dnn_detect_post_proc(AVFrame *frame, DNNData *output, AVFilterContext *filter_ctx)
+{
+    DnnDetectContext *ctx = filter_ctx->priv;
+    float conf_threshold = ctx->conf;
+    int proposal_count = output->height;
+    int detect_size = output->width;
+    float *detections = output->data;
+    int nb_bbox = 0;
+    AVFrameSideData *sd;
+    AVDnnBoundingBox *bbox;
+
+    sd = av_frame_get_side_data(frame, AV_FRAME_DATA_DNN_BBOXES);
+    if (sd) {
+        av_log(filter_ctx, AV_LOG_ERROR, "already have dnn bounding boxes in side data.\n");
+        return -1;
+    }
+
+    for (int i = 0; i < proposal_count; ++i) {
+            float conf = detections[i * detect_size + 2];
+            if (conf < conf_threshold) {
+                continue;
+            }
+            nb_bbox++;
+    }
+
+    if (nb_bbox == 0) {
+        av_log(filter_ctx, AV_LOG_VERBOSE, "nothing detected in this frame.\n");
+        return 0;
+    }
+
+    sd = av_frame_new_side_data(frame, AV_FRAME_DATA_DNN_BBOXES,
+                                sizeof(AVDnnBoundingBox) * nb_bbox);
+    if (!sd) {
+        av_log(filter_ctx, AV_LOG_ERROR, "failed to allocate side data for AV_FRAME_DATA_DNN_BBOXES with %d bboxes\n", nb_bbox);
+        return -1;
+    }
+
+    bbox = (AVDnnBoundingBox *)sd->data;
+    for (int i = 0; i < proposal_count; ++i) {
+            int av_unused image_id = (int)detections[i * detect_size + 0];
+            int label = (int)detections[i * detect_size + 1];
+            float conf = detections[i * detect_size + 2];
+            float x0 = detections[i * detect_size + 3];
+            float y0 = detections[i * detect_size + 4];
+            float x1 = detections[i * detect_size + 5];
+            float y1 = detections[i * detect_size + 6];
+
+            if (conf < conf_threshold) {
+                continue;
+            }
+
+            *bbox = (AVDnnBoundingBox) {
+                .self_size = sizeof(*bbox),
+                .model_input_width = ctx->model_input_width,
+                .model_input_height = ctx->model_input_height,
+                .left = (int)(x0 * ctx->model_input_width),
+                .right = (int)(x1 * ctx->model_input_width),
+                .top = (int)(y0 * ctx->model_input_height),
+                .bottom = (int)(y1 * ctx->model_input_height),
+                .detect_label = label,
+                .detect_conf = av_make_q((int)(conf * 10000), 10000),
+                .classify_count = 0,
+            };
+
+            nb_bbox--;
+            if (nb_bbox == 0) {
+                break;
+            }
+            bbox++;
+    }
+
+    return 0;
+}
+
+static av_cold int dnn_detect_init(AVFilterContext *context)
+{
+    DNNReturnType result;
+    DNNData model_input;
+    DnnDetectContext *ctx = context->priv;
+    int ret = ff_dnn_init(&ctx->dnnctx, DFT_ANALYTICS_DETECT, context);
+    if (ret < 0)
+        return ret;
+    ff_dnn_set_proc(&ctx->dnnctx, NULL, dnn_detect_post_proc);
+
+    result = ff_dnn_get_input(&ctx->dnnctx, &model_input);
+    if (result != DNN_SUCCESS) {
+        av_log(ctx, AV_LOG_ERROR, "could not get input from the model.\n");
+        return AVERROR(EIO);
+    }
+
+    ctx->model_input_width = model_input.width;
+    ctx->model_input_height = model_input.height;
+    return 0;
+}
+
+static int dnn_detect_query_formats(AVFilterContext *context)
+{
+    static const enum AVPixelFormat pix_fmts[] = {
+        AV_PIX_FMT_RGB24, AV_PIX_FMT_BGR24,
+        AV_PIX_FMT_GRAY8, AV_PIX_FMT_GRAYF32,
+        AV_PIX_FMT_YUV420P, AV_PIX_FMT_YUV422P,
+        AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV410P, AV_PIX_FMT_YUV411P,
+        AV_PIX_FMT_NV12,
+        AV_PIX_FMT_NONE
+    };
+    AVFilterFormats *fmts_list = ff_make_format_list(pix_fmts);
+    return ff_set_common_formats(context, fmts_list);
+}
+
+static int dnn_detect_filter_frame(AVFilterLink *inlink, AVFrame *in)
+{
+    AVFilterContext *context  = inlink->dst;
+    AVFilterLink *outlink = context->outputs[0];
+    DnnDetectContext *ctx = context->priv;
+    DNNReturnType dnn_result;
+
+    dnn_result = ff_dnn_execute_model(&ctx->dnnctx, in, NULL);
+    if (dnn_result != DNN_SUCCESS){
+        av_log(ctx, AV_LOG_ERROR, "failed to execute model\n");
+        av_frame_free(&in);
+        return AVERROR(EIO);
+    }
+
+    return ff_filter_frame(outlink, in);
+}
+
+static int dnn_detect_activate_sync(AVFilterContext *filter_ctx)
+{
+    AVFilterLink *inlink = filter_ctx->inputs[0];
+    AVFilterLink *outlink = filter_ctx->outputs[0];
+    AVFrame *in = NULL;
+    int64_t pts;
+    int ret, status;
+    int got_frame = 0;
+
+    FF_FILTER_FORWARD_STATUS_BACK(outlink, inlink);
+
+    do {
+        // drain all input frames
+        ret = ff_inlink_consume_frame(inlink, &in);
+        if (ret < 0)
+            return ret;
+        if (ret > 0) {
+            ret = dnn_detect_filter_frame(inlink, in);
+            if (ret < 0)
+                return ret;
+            got_frame = 1;
+        }
+    } while (ret > 0);
+
+    // if frame got, schedule to next filter
+    if (got_frame)
+        return 0;
+
+    if (ff_inlink_acknowledge_status(inlink, &status, &pts)) {
+        if (status == AVERROR_EOF) {
+            ff_outlink_set_status(outlink, status, pts);
+            return ret;
+        }
+    }
+
+    FF_FILTER_FORWARD_WANTED(outlink, inlink);
+
+    return FFERROR_NOT_READY;
+}
+
+static int dnn_detect_flush_frame(AVFilterLink *outlink, int64_t pts, int64_t *out_pts)
+{
+    DnnDetectContext *ctx = outlink->src->priv;
+    int ret;
+    DNNAsyncStatusType async_state;
+
+    ret = ff_dnn_flush(&ctx->dnnctx);
+    if (ret != DNN_SUCCESS) {
+        return -1;
+    }
+
+    do {
+        AVFrame *in_frame = NULL;
+        AVFrame *out_frame = NULL;
+        async_state = ff_dnn_get_async_result(&ctx->dnnctx, &in_frame, &out_frame);
+        if (out_frame) {
+            av_assert0(in_frame == out_frame);
+            ret = ff_filter_frame(outlink, out_frame);
+            if (ret < 0)
+                return ret;
+            if (out_pts)
+                *out_pts = out_frame->pts + pts;
+        }
+        av_usleep(5000);
+    } while (async_state >= DAST_NOT_READY);
+
+    return 0;
+}
+
+static int dnn_detect_activate_async(AVFilterContext *filter_ctx)
+{
+    AVFilterLink *inlink = filter_ctx->inputs[0];
+    AVFilterLink *outlink = filter_ctx->outputs[0];
+    DnnDetectContext *ctx = filter_ctx->priv;
+    AVFrame *in = NULL;
+    int64_t pts;
+    int ret, status;
+    int got_frame = 0;
+    int async_state;
+
+    FF_FILTER_FORWARD_STATUS_BACK(outlink, inlink);
+
+    do {
+        // drain all input frames
+        ret = ff_inlink_consume_frame(inlink, &in);
+        if (ret < 0)
+            return ret;
+        if (ret > 0) {
+            if (ff_dnn_execute_model_async(&ctx->dnnctx, in, NULL) != DNN_SUCCESS) {
+                return AVERROR(EIO);
+            }
+        }
+    } while (ret > 0);
+
+    // drain all processed frames
+    do {
+        AVFrame *in_frame = NULL;
+        AVFrame *out_frame = NULL;
+        async_state = ff_dnn_get_async_result(&ctx->dnnctx, &in_frame, &out_frame);
+        if (out_frame) {
+            av_assert0(in_frame == out_frame);
+            ret = ff_filter_frame(outlink, out_frame);
+            if (ret < 0)
+                return ret;
+            got_frame = 1;
+        }
+    } while (async_state == DAST_SUCCESS);
+
+    // if frame got, schedule to next filter
+    if (got_frame)
+        return 0;
+
+    if (ff_inlink_acknowledge_status(inlink, &status, &pts)) {
+        if (status == AVERROR_EOF) {
+            int64_t out_pts = pts;
+            ret = dnn_detect_flush_frame(outlink, pts, &out_pts);
+            ff_outlink_set_status(outlink, status, out_pts);
+            return ret;
+        }
+    }
+
+    FF_FILTER_FORWARD_WANTED(outlink, inlink);
+
+    return 0;
+}
+
+static int dnn_detect_activate(AVFilterContext *filter_ctx)
+{
+    DnnDetectContext *ctx = filter_ctx->priv;
+
+    if (ctx->dnnctx.async)
+        return dnn_detect_activate_async(filter_ctx);
+    else
+        return dnn_detect_activate_sync(filter_ctx);
+}
+
+static av_cold void dnn_detect_uninit(AVFilterContext *ctx)
+{
+    DnnDetectContext *context = ctx->priv;
+    ff_dnn_uninit(&context->dnnctx);
+}
+
+static const AVFilterPad dnn_detect_inputs[] = {
+    {
+        .name         = "default",
+        .type         = AVMEDIA_TYPE_VIDEO,
+    },
+    { NULL }
+};
+
+static const AVFilterPad dnn_detect_outputs[] = {
+    {
+        .name = "default",
+        .type = AVMEDIA_TYPE_VIDEO,
+    },
+    { NULL }
+};
+
+AVFilter ff_vf_dnn_detect = {
+    .name          = "dnn_detect",
+    .description   = NULL_IF_CONFIG_SMALL("Apply DNN detect filter to the input."),
+    .priv_size     = sizeof(DnnDetectContext),
+    .init          = dnn_detect_init,
+    .uninit        = dnn_detect_uninit,
+    .query_formats = dnn_detect_query_formats,
+    .inputs        = dnn_detect_inputs,
+    .outputs       = dnn_detect_outputs,
+    .priv_class    = &dnn_detect_class,
+    .activate      = dnn_detect_activate,
+};
-- 
2.17.1



More information about the ffmpeg-devel mailing list