ntua-edge-robotics/yoloaccuracy

Benchmark the accuracy of YOLOv5 using libjxl.

YoloAccuracy

The project aims to benchmark the accuracy of YOLOv5 using libjxl.

Read, watch or cite our paper

Installation

1. Install libjxl on your computer. cjxl and djxl needs to be available on your path.
2. Install the dependencies with pip install -r requirements.txt

Primary Data Generation

The API of the script can be found using python src/generate_primary_data.py -h. The script follows these steps :

1. Load a raw or lossless image (e.g. PNG)
2. Resize the image (if necessary) and keep the aspect ratio
3. Save the image as JPEG XL (JXL) according to a quality factor
4. Convert the JXL to a lossless format (PNG)
5. Detect the objects in the saved JXL with YOLOv5
6. Get the inference speed, the predicted classes, the confidence (%) and the actual classes
7. Save the results in a JSON file

Notes

The script is meant to work with the People Overhead dataset. To use another dataset, dataset_utils.py should be updated to retrieve the proper actual classes.

Accuracy Visualization

The following script will produce a graphic of the predictions accuracy according to the height and the quality factor. The API of the script can be found using python src/visualize_accuracy.py -h.

What could be improved

• Log and handle errors
• Automated tests