yolov4_tiny_rs is a rust implementation of the yolov4_tiny algorithm using nvidia's cuda libraries.

This is a minimal implementation which only preforms inference.

The original paper for YOLOv4 can be found here: https://arxiv.org/abs/2004.10934 The author's personal site can be found here: https://pjreddie.com/darknet/yolo/

Setup

To compile and run Nvidia's cuda, cutensor, cudnn, and cublas libraries must be installed. This program assumes the use of CUDA 11.4. On windows it is expected the relevant binaries to be found in C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v11.4\\lib\\x64\\ . On linux /usr/local/cuda/include . If you have installed the cuda dependencies in a different location you will need to update the cuda rs build files accordingly.

An external set of weights is needed to preform inference. The weights can be found here. Once downloaded place the weights yolov4_tiny_tg in the data folder.

Run

NOTE: This program requires the user to have a nvidia gpu with a compute core rating over 7. To run cargo run (--release) <input_file_name> -o <optional_output_file_name>. The output file will render any bounding boxes, class names and confidence scores to a copy of the original image.

TODO

• user defined score criteria
• implement macOs implementation
• implement a sensible automatic deallocation systems, probably ref-counting.