Hi,
We managed to update the deepstream-yolo-app in accordance with deepstream-test3-app to run multiple streams with yolo. When we run the deepstream-test3-app with the resnet10 architecture the detections for each stream are reliable. However, when we run the deepstream-yolo-app with multiple streams ( usually > 4) , the detections become horrible i.e only 1 or 2 objects get detected if streams are increased above 4. Is this because the Yolo architecture is more complicated than the resnet10 arch or are we missing out on something?
Hi,
Can you please provide more details regarding your setup environment ? What GPU are you using ? which yolo model are you running inference with ? What changes have you done in the app ? What kind of sources are you using ? Can you share a picture of the missing detections ?
Hi @NvCJR,
We are using the Yolov3 architecture and the GPU is a 2080Ti. All are sources are h.264 videos currently but we will shift to rtsp.Our deepstream-yolo-app.c is as follows:
/*
* Copyright (c) 2018 NVIDIA Corporation. All rights reserved.
*
* NVIDIA Corporation and its licensors retain all intellectual property
* and proprietary rights in and to this software, related documentation
* and any modifications thereto. Any use, reproduction, disclosure or
* distribution of this software and related documentation without an express
* license agreement from NVIDIA Corporation is strictly prohibited.
*
*/
//#include <iostream>
#include <gst/gst.h>
#include <glib.h>
#include <math.h>
#include <string.h>
#include <sys/time.h>
#include <stdio.h>
#include "gstnvdsmeta.h"
#include "gstnvstreammeta.h"
#include "gst-nvmessage.h"
#include<vector>
#include<string>
#include<iostream>
#define MAX_DISPLAY_LEN 1024
#define PGIE_CLASS_ID_VEHICLE 0
#define YOLO_UNIQUE_ID 15
#define PGIE_CLASS_ID_PERSON 2
gint frame_number = 0;
/* The muxer output resolution must be set if the input streams will be of
* different resolution. The muxer will scale all the input frames to this
* resolution. */
#define MUXER_OUTPUT_WIDTH 1920
#define MUXER_OUTPUT_HEIGHT 1080
/* Muxer batch formation timeout, for e.g. 40 millisec. Should ideally be set
* based on the fastest source's framerate. */
#define MUXER_BATCH_TIMEOUT_USEC 4000000
#define TILED_OUTPUT_WIDTH 1920
#define TILED_OUTPUT_HEIGHT 1080
/* NVIDIA Decoder source pad memory feature. This feature signifies that source
* pads having this capability will push GstBuffers containing cuda buffers. */
#define GST_CAPS_FEATURES_NVMM "memory:NVMM"
gchar pgie_classes_str[4][32] = { "Vehicle", "TwoWheeler", "Person",
"RoadSign"
};
#define FPS_PRINT_INTERVAL 10
static struct timeval start_time = { };
static guint probe_counter = 0;
/* tiler_sink_pad_buffer_probe will extract metadata received on OSD sink pad
* and update params for drawing rectangle, object information etc. */
static GstPadProbeReturn
tiler_src_pad_buffer_probe (GstPad * pad, GstPadProbeInfo * info,
gpointer u_data)
{
GstMeta *gst_meta = NULL;
NvDsMeta *nvdsmeta = NULL;
gpointer state = NULL;
static GQuark _nvdsmeta_quark = 0;
GstBuffer *buf = (GstBuffer *) info->data;
NvDsFrameMeta *frame_meta = NULL;
guint num_rects = 0, rect_index = 0;
NvDsObjectParams *obj_meta = NULL;
NvOSD_TextParams *txt_params = NULL;
GstNvStreamMeta *streammeta = NULL;
if (!_nvdsmeta_quark)
_nvdsmeta_quark = g_quark_from_static_string (NVDS_META_STRING);
if (probe_counter == 0) {
gettimeofday (&start_time, NULL);
} else if (probe_counter == FPS_PRINT_INTERVAL) {
struct timeval cur_time;
gettimeofday (&cur_time, NULL);
g_print ("FPS for the last %d batches: %.2f\n", FPS_PRINT_INTERVAL,
FPS_PRINT_INTERVAL / ((cur_time.tv_sec - start_time.tv_sec) +
(cur_time.tv_usec - start_time.tv_usec) / 1e6));
probe_counter = 0;
start_time = cur_time;
}
probe_counter++;
streammeta = gst_buffer_get_nvstream_meta (buf);
while ((gst_meta = gst_buffer_iterate_meta (buf, &state))) {
if (gst_meta_api_type_has_tag (gst_meta->info->api, _nvdsmeta_quark)) {
nvdsmeta = (NvDsMeta *) gst_meta;
/* We are interested only in intercepting Meta of type
* "NVDS_META_FRAME_INFO" as they are from our infer elements. */
if (nvdsmeta->meta_type == NVDS_META_FRAME_INFO) {
guint vehicle_count = 0;
guint person_count = 0;
frame_meta = (NvDsFrameMeta *) nvdsmeta->meta_data;
if (frame_meta == NULL) {
g_print ("NvDS Meta contained NULL meta \n");
return GST_PAD_PROBE_OK;
}
/* We reset the num_strings here as we plan to iterate through the
* the detected objects and form our own strings.
* The pipeline generated strings shall be discarded.
*/
frame_meta->num_strings = 0;
num_rects = frame_meta->num_rects;
/* This means we have num_rects in frame_meta->obj_params,
* now lets iterate through them */
for (rect_index = 0; rect_index < num_rects; rect_index++) {
/* Now using above information we need to form a text that should
* be displayed on top of the bounding box, so lets form it here. */
obj_meta = (NvDsObjectParams *) & frame_meta->obj_params[rect_index];
txt_params = &(obj_meta->text_params);
if (txt_params->display_text)
g_free (txt_params->display_text);
txt_params->display_text = (char*)g_malloc0 (MAX_DISPLAY_LEN);
g_snprintf (txt_params->display_text, MAX_DISPLAY_LEN, "%s ",
pgie_classes_str[obj_meta->class_id]);
if (obj_meta->class_id == PGIE_CLASS_ID_VEHICLE)
vehicle_count++;
if (obj_meta->class_id == PGIE_CLASS_ID_PERSON)
person_count++;
/* Now set the offsets where the string should appear */
txt_params->x_offset = obj_meta->rect_params.left;
txt_params->y_offset = obj_meta->rect_params.top - 25;
/* Font , font-color and font-size */
txt_params->font_params.font_name = "Arial";
txt_params->font_params.font_size = 10;
txt_params->font_params.font_color.red = 1.0;
txt_params->font_params.font_color.green = 1.0;
txt_params->font_params.font_color.blue = 1.0;
txt_params->font_params.font_color.alpha = 1.0;
/* Text background color */
txt_params->set_bg_clr = 1;
txt_params->text_bg_clr.red = 0.0;
txt_params->text_bg_clr.green = 0.0;
txt_params->text_bg_clr.blue = 0.0;
txt_params->text_bg_clr.alpha = 1.0;
frame_meta->num_strings++;
}
/* Enable this print by setting environment variable GST_DEBUG=4
* before running the app. */
g_print ("Source %d Frame Number = %lu Number of objects = %d "
"Vehicle Count = %d Person Count = %d", frame_meta->stream_id,
streammeta->stream_frame_num[frame_meta->batch_id],
num_rects, vehicle_count, person_count);
}
}
}
return GST_PAD_PROBE_OK;
}
static gboolean
bus_call (GstBus * bus, GstMessage * msg, gpointer data)
{
GMainLoop *loop = (GMainLoop *) data;
switch (GST_MESSAGE_TYPE (msg)) {
case GST_MESSAGE_EOS:
g_print ("End of stream\n");
g_main_loop_quit (loop);
break;
case GST_MESSAGE_WARNING:
{
gchar *debug;
GError *error;
gst_message_parse_warning (msg, &error, &debug);
g_printerr ("WARNING from element %s: %s\n",
GST_OBJECT_NAME (msg->src), error->message);
g_free (debug);
g_printerr ("Warning: %s\n", error->message);
g_error_free (error);
break;
}
case GST_MESSAGE_ERROR:
{
gchar *debug;
GError *error;
gst_message_parse_error (msg, &error, &debug);
g_printerr ("ERROR from element %s: %s\n",
GST_OBJECT_NAME (msg->src), error->message);
if (debug)
g_printerr ("Error details: %s\n", debug);
g_free (debug);
g_error_free (error);
g_main_loop_quit (loop);
break;
}
case GST_MESSAGE_ELEMENT:
{
if (gst_nvmessage_is_stream_eos (msg)) {
guint stream_id;
if (gst_nvmessage_parse_stream_eos (msg, &stream_id)) {
g_print ("Got EOS from stream %d\n", stream_id);
}
}
break;
}
default:
break;
}
return TRUE;
}
static void
cb_newpad (GstElement * decodebin, GstPad * decoder_src_pad, gpointer data)
{
GstCaps *caps = gst_pad_query_caps (decoder_src_pad, NULL);
const GstStructure *str = gst_caps_get_structure (caps, 0);
const gchar *name = gst_structure_get_name (str);
GstElement *source_bin = (GstElement *) data;
GstCapsFeatures *features = gst_caps_get_features (caps, 0);
/* Need to check if the pad created by the decodebin is for video and not
* audio. */
if (!strncmp (name, "video", 5)) {
/* Link the decodebin pad only if decodebin has picked nvidia
* decoder plugin nvdec_*. We do this by checking if the pad caps contain
* NVMM memory features. */
if (gst_caps_features_contains (features, GST_CAPS_FEATURES_NVMM)) {
/* Get the source bin ghost pad */
GstPad *bin_ghost_pad = gst_element_get_static_pad (source_bin, "src");
if (!gst_ghost_pad_set_target (GST_GHOST_PAD (bin_ghost_pad),
decoder_src_pad)) {
g_printerr ("Failed to link decoder src pad to source bin ghost pad\n");
}
gst_object_unref (bin_ghost_pad);
} else {
g_printerr ("Error: Decodebin did not pick nvidia decoder plugin.\n");
}
}
}
static GstElement *
create_source_bin (guint index, gchar * uri)
{
GstElement *bin = NULL, *uri_decode_bin = NULL;
gchar bin_name[16] = { };
g_snprintf (bin_name, 15, "source-bin-%02d", index);
/* Create a source GstBin to abstract this bin's content from the rest of the
* pipeline */
bin = gst_bin_new (bin_name);
/* Source element for reading from the uri.
* We will use decodebin and let it figure out the container format of the
* stream and the codec and plug the appropriate demux and decode plugins. */
uri_decode_bin = gst_element_factory_make ("uridecodebin", "uri-decode-bin");
if (!bin || !uri_decode_bin) {
g_printerr ("One element in source bin could not be created.\n");
return NULL;
}
/* We set the input uri to the source element */
g_object_set (G_OBJECT (uri_decode_bin), "uri", uri, NULL);
/* Connect to the "pad-added" signal of the decodebin which generates a
* callback once a new pad for raw data has beed created by the decodebin */
g_signal_connect (G_OBJECT (uri_decode_bin), "pad-added",
G_CALLBACK (cb_newpad), bin);
gst_bin_add (GST_BIN (bin), uri_decode_bin);
/* We need to create a ghost pad for the source bin which will act as a proxy
* for the video decoder src pad. The ghost pad will not have a target right
* now. Once the decode bin creates the video decoder and generates the
* cb_newpad callback, we will set the ghost pad target to the video decoder
* src pad. */
if (!gst_element_add_pad (bin, gst_ghost_pad_new_no_target ("src",
GST_PAD_SRC))) {
g_printerr ("Failed to add ghost pad in source bin\n");
return NULL;
}
return bin;
}
int
main (int argc, char *argv[])
{
GMainLoop *loop = NULL;
GstElement *pipeline = NULL, *streammux = NULL, *sink = NULL, *pgie = NULL,
*nvvidconv = NULL, *nvosd = NULL, *tiler = NULL, *filter1=NULL, *filter2=NULL;
GstBus *bus = NULL;
guint bus_watch_id;
GstPad *tiler_src_pad = NULL;
guint i, num_sources;
guint tiler_rows, tiler_columns;
guint pgie_batch_size;
GstCaps *caps1 = NULL, *caps2 = NULL;
/* Check input arguments */
if (argc < 2) {
g_printerr ("Usage: %s <uri1> [uri2] ... [uriN] \n", argv[0]);
return -1;
}
num_sources=argc - 1;
/* Standard GStreamer initialization */
gst_init (&argc, &argv);
loop = g_main_loop_new (NULL, FALSE);
/* Create gstreamer elements */
/* Create Pipeline element that will form a connection of other elements */
pipeline = gst_pipeline_new ("dstest3-pipeline");
/* Create nvstreammux instance to form batches from one or more sources. */
streammux = gst_element_factory_make ("nvstreammux", "stream-muxer");
if (!pipeline || !streammux) {
g_printerr ("One element could not be created. Exiting.\n");
return -1;
}
gst_bin_add (GST_BIN (pipeline), streammux);
for (i = 0; i < num_sources; i++) {
GstPad *sinkpad, *srcpad;
gchar pad_name[16] = { };
GstElement *source_bin = create_source_bin (i, argv[i + 1]);
if (!source_bin) {
g_printerr ("Failed to create source bin. Exiting.\n");
return -1;
}
gst_bin_add (GST_BIN (pipeline), source_bin);
g_snprintf (pad_name, 15, "sink_%u", i);
sinkpad = gst_element_get_request_pad (streammux, pad_name);
if (!sinkpad) {
g_printerr ("Streammux request sink pad failed. Exiting.\n");
return -1;
}
srcpad = gst_element_get_static_pad (source_bin, "src");
if (!srcpad) {
g_printerr ("Failed to get src pad of source bin. Exiting.\n");
return -1;
}
if (gst_pad_link (srcpad, sinkpad) != GST_PAD_LINK_OK) {
g_printerr ("Failed to link source bin to stream muxer. Exiting.\n");
return -1;
}
gst_object_unref (srcpad);
gst_object_unref (sinkpad);
}
filter1 = gst_element_factory_make ("capsfilter", "filter1");
filter2 = gst_element_factory_make ("capsfilter", "filter2");
/* Use nvinfer to infer on batched frame. */
pgie = gst_element_factory_make ("nvyolo", "yolo-inference-engine");
/* Use nvtiler to composite the batched frames into a 2D tiled array based
* on the source of the frames. */
tiler = gst_element_factory_make ("nvmultistreamtiler", "nvtiler");
/* Use convertor to convert from NV12 to RGBA as required by nvosd */
nvvidconv = gst_element_factory_make ("nvvidconv", "nvvideo-converter");
/* Create OSD to draw on the converted RGBA buffer */
nvosd = gst_element_factory_make ("nvosd", "nv-onscreendisplay");
/* Finally render the osd output nveglglessink */
sink = gst_element_factory_make ("fakesink", "nvvideo-renderer");
if (!pgie || !tiler || !nvvidconv || !nvosd || !sink) {
g_printerr ("One element could not be created. Exiting.\n");
return -1;
}
g_object_set (G_OBJECT (streammux), "width", MUXER_OUTPUT_WIDTH, "height",
MUXER_OUTPUT_HEIGHT, "batch-size", num_sources,
"batched-push-timeout", MUXER_BATCH_TIMEOUT_USEC, NULL);
/* Configure the nvinfer element using the nvinfer config file. */
g_object_set (G_OBJECT (pgie),
"config-file-path", "/home/rbcgpu2/Documents/deepstream/deepstream_reference_apps/yolo/config/yolov3.txt", NULL);
/*
g_object_get (G_OBJECT (pgie), "batch-size", &pgie_batch_size, NULL);
if (pgie_batch_size != num_sources) {
g_printerr
("WARNING: Overriding infer-config batch-size (%d) with number of sources (%d)\n",
pgie_batch_size, num_sources);
g_object_set (G_OBJECT (pgie), "batch-size", num_sources, NULL);
}*/
/* Override the batch-size set in the config file with the number of sources. */
tiler_rows = (guint) sqrt (num_sources);
tiler_columns = (guint) ceil (1.0 * num_sources / tiler_rows);
/* we set the tiler properties here */
g_object_set (G_OBJECT (tiler), "rows", tiler_rows, "columns", tiler_columns,
"width", TILED_OUTPUT_WIDTH, "height", TILED_OUTPUT_HEIGHT, NULL);
/* we set the osd properties here */
g_object_set (G_OBJECT (nvosd), "font-size", 15, NULL);
/* we add a message handler */
bus = gst_pipeline_get_bus (GST_PIPELINE (pipeline));
bus_watch_id = gst_bus_add_watch (bus, bus_call, loop);
gst_object_unref (bus);
/* Set up the pipeline */
/* we add all elements into the pipeline */
gst_bin_add_many (GST_BIN (pipeline), tiler, filter1, nvvidconv, filter2, pgie,nvosd, sink,
NULL);
caps1 = gst_caps_from_string ("video/x-raw(memory:NVMM), format=NV12");
g_object_set (G_OBJECT (filter1), "caps", caps1, NULL);
gst_caps_unref (caps1);
caps2 = gst_caps_from_string ("video/x-raw(memory:NVMM), format=RGBA");
g_object_set (G_OBJECT (filter2), "caps", caps2, NULL);
gst_caps_unref (caps2);
/* we link the elements together
* nvstreammux -> nvinfer -> nvtiler -> nvvidconv -> nvosd -> video-renderer */
if (!gst_element_link_many (streammux, tiler, filter1, nvvidconv, filter2, pgie, nvosd, sink,
NULL)) {
g_printerr ("Elements could not be linked. Exiting.\n");
return -1;
}
/* Lets add probe to get informed of the meta data generated, we add probe to
* the sink pad of the osd element, since by that time, the buffer would have
* had got all the metadata. */
tiler_src_pad = gst_element_get_static_pad (pgie, "src");
if (!tiler_src_pad)
g_print ("Unable to get src pad\n");
else
gst_pad_add_probe (tiler_src_pad, GST_PAD_PROBE_TYPE_BUFFER,
tiler_src_pad_buffer_probe, NULL, NULL);
/* Set the pipeline to "playing" state */
g_print ("Now playing:");
for (i = 0; i < num_sources; i++) {
g_print (" %s,", argv[i + 1]);
}
g_print ("\n");
gst_element_set_state (pipeline, GST_STATE_PLAYING);
/* Wait till pipeline encounters an error or EOS */
g_print ("Running...\n");
g_main_loop_run (loop);
/* Out of the main loop, clean up nicely */
g_print ("Returned, stopping playback\n");
gst_element_set_state (pipeline, GST_STATE_NULL);
g_print ("Deleting pipeline\n");
gst_object_unref (GST_OBJECT (pipeline));
g_source_remove (bus_watch_id);
g_main_loop_unref (loop);
return 0;
}
Also, i wanted to know if we could assign particular GPU memory for each stream becuase maybe the detections will get better with this. I cannot share any pictures right now but detections are completely absent when more than 2 streams are used i.e:
If I play the same video 4 times, the first 2 windows have proper dectection but the remaining windows show no detections at all.
Can you also post the logs you see on screen when the app is launched ?
yolov3 is a compute intensive model, you can try switching it with yolov3-tiny and see if it helps.
Hi @NvCJR,
The logs aren’t usable becuase we haven’t fixed the g_print to alternate between sources and show the detection of each source. But we clearly can see the detections reduce significantly when we increase the number of streams. I agree Yolov3 is compute intensive but we have to stick with this model because of several other applications that are dependent on it. However, why isn’t deepstream-yolo-app making use of our entire GPU memory. Don’t you think the detections will get better if we could use the entire memory instead of the 3GB/12GB that it is currently using? Is there anyway I can make the app use all the 12GB available?
Memory usage is not the only metric that defines performance. It also depends on your gpu utilization. You can try running higher number of streams at a lower frame rate in your present setup. As a work around, you can launch multiple pipelines in parallel, however the inference throughput may still be affected depending on the GPU utilization. You can also try inference with INT8 precision for better throughput as well.