This project provides YOLOv3 and YOLOv4 implementations for object detection with support for multiple datasets including COCO, BDD100k, and Visdrone. It implements model compression through pruning (channel and layer), quantization-aware training (8-bit), and knowledge distillation. Multiple backbone architectures are supported: standard Darknet, Tiny-YOLO, and MobileNetV3-based variants for different speed/accuracy tradeoffs. The repository enables multi-dataset training and provides pre-trained weights for various configurations.