yolov5-crowdhuman-onnx

YOLOv5 CrowdHuman ONNX Runtime

Video by Coverr from Pexels: https://www.pexels.com/video/black-and-white-video-of-people-853889/

This repository contains code and instructions for performing object detection using the YOLOv5 model with the CrowdHuman dataset, utilizing ONNX Runtime for inference.

Features

• Inference using ONNX Runtime with GPU (tested on Ubuntu).
• Easy-to-use Python scripts for inference.
• Supports multiple input formats: image, video, or webcam.

Installation

Clone the Repository

git clone https://github.com/yourusername/yolov5-crowdhuman-onnx.git
cd yolov5-crowdhuman-onnx

Install Required Packages

pip install -r requirements.txt

Usage

Before running inference, you need to download weights of the YOLOv5m model trained on CrowdHuman dataset in ONNX format.

Download weights (Linux)

sh download.sh

Download weights from the following links

Note: The weights are saved in FP32.

Model Name	ONNX Model Link	Number of Parameters	Model Size
YOLOv5m	crowdhuman.onnx	21.2M	84 MB

If you have custom weights, you can convert your weights to ONNX format. Follow the instructions in the YOLOv5 repository to convert your model. You can use the converted ONNX model with this repository.

Inference

python main.py --weights weights/crowdhuman.onnx --source assets/vid_input.mp4 # video
                                                 --source 0 --view # webcam and display
                                                 --source assets/img_input.jpg # image

• To save results add the --save argument and results will be saved under the runs folder
• To display video add the --view argument

Command Line Arguments

usage: main.py [-h] [--weights WEIGHTS] [--source SOURCE] [--img-size IMG_SIZE [IMG_SIZE ...]] [--conf-thres CONF_THRES] [--iou-thres IOU_THRES]
               [--max-det MAX_DET] [--save] [--view] [--project PROJECT] [--name NAME]

options:
  -h, --help            show this help message and exit
  --weights WEIGHTS     model path
  --source SOURCE       Path to video/image/webcam
  --img-size IMG_SIZE [IMG_SIZE ...]
                        inference size h,w
  --conf-thres CONF_THRES
                        confidence threshold
  --iou-thres IOU_THRES
                        NMS IoU threshold
  --max-det MAX_DET     maximum detections per image
  --save                Save detected images
  --view                View inferenced images
  --project PROJECT     save results to project/name
  --name NAME           save results to project/name

Reference

1. https://github.com/ultralytics/yolov5
2. Thanks for the model weight to SibiAkkash