Deepstream real-time instance segmentation

Real-time instance segmentation

Youtube link:

Support model

• YOLOv5-Seg
• YOLOv7-Seg

Prerequisites

• Deepstream 6.0.1
• TensorRT-OSS: github.com/hiennguyen9874/TensorRT

Getting started

Convert model to engine file

• Download model from github.com/hiennguyen9874/deepstream-yolov7-mask/releases/tag/v0.1 or export onnx file from weight:

  • github.com/hiennguyen9874/yolov5-seg
  • github.com/hiennguyen9874/yolov7-seg

• Export onnx to tensorRT: /usr/src/tensorrt/bin/trtexec --onnx=./models/yolov7-seg-dev.onnx --saveEngine=./models/yolov7-seg-dev.trt --fp16 --workspace=8192

Build custom parser

• cd nvdsinfer_customparser
• make

Run

• bash ./run.sh

Docker

Get cuda version number

• docker run -v ${PWD}:/app/ -w /app --rm --gpus all -it nvcr.io/nvidia/deepstream:6.0.1-devel bash -c 'nvcc deviceQuery.cpp -o deviceQuery && ./deviceQuery'

• Modify DGPU_ARCHS in docker compose

• Example (DGPU_ARCHS=75):

  

Run

• build: docker-compose build

• up: docker-compose up

• Output video in outputs folder

Benchmark

Test on T4

Model	FPS	mAPval 0.5:0.95 (Box)	mAPval 0.5:0.95 (Mask)
YOLOv5n	247.97	0.27	0.234
YOLOv5s	183.77	0.372	0.317
YOLOv5m	119.58	0.445	0.373
YOLOv5l	85.80	0.484	0.401
YOLOv5x	54.04	0.502	0.415
YOLOv7-Seg	84.12	0.49	0.406
YOLOv7-Seg-Dev	84.59	0.505	0.414
YOLOv7x-Seg	60.29	0.511	0.421