Object Detection Across Image Tiles

Introduction

This application uses YoloV8 to detect objects within images across tiles. Total runtime is printed for comparison between different model architectures.

Optimization

• Model can be converted to ONNX format for faster inference speed on CPU/GPU.
• Implemented parallel computing to optimize postprocessing on multiple tiles.
• Introduced a buffer around each image to enhance detection accuracy.
• Set up a Docker environment to provide the app as a standalone executable.

Usage

Basic Usage

Install requirements

pip install -r requirements.txt

To run with pytorch. Default GPU, fallback to CPU.

python main.py --input_path "your/input/path/" --output_path "your/output/path/"

To run with ONNX using CPU only

python main.py --input_path "your/input/path/" --output_path "your/output/path/ --use_onnx"

To run with ONNX using GPU, fallback to CPU

python main.py --input_path "your/input/path/" --output_path "your/output/path/ --use_onnx --onnx_gpu"

Sample command

Use PyTorch, runs on GPU, fallback to CPU

python main.py -i test/ --model_name yolov8x --class_yaml coco8.yaml

Use onnx, runs on CPU

python main.py -i test/ --model_name yolov8x --class_yaml coco8.yaml --use_onnx

Use onnx, runs on GPU, fallback to CPU

python main.py -i test/ --model_name yolov8x --class_yaml coco8.yaml --use_onnx --onnx_gpu