This repository delivers a comprehensive, modular, and research-friendly implementation of YOLOv2 (You Only Look Once, version 2) in PyTorch, powered by the Darknet-19 backbone.
YOLOv2 marked a turning point in real-time object detection, balancing speed and accuracy in ways that reshaped the field. With its introduction of anchor boxes, passthrough connections, and a streamlined architecture, YOLOv2 became a foundation upon which later detectors (YOLOv3, YOLOv4, and beyond) were built.
Our implementation aims to capture not only the original design but also to present it in a clean, transparent, and fully modular structure, making it accessible for:
- Researchers, who want to experiment with YOLOv2 as a baseline or extend it to new ideas.
- Students and educators, seeking a readable and didactic PyTorch reimplementation to understand the internals of modern detectors.
- Practitioners, who may want to adapt YOLOv2 to custom datasets, test mAP performance, or benchmark lightweight models.
The repository covers every stage of the detection pipeline, from data preparation to evaluation and visualization:
- Dataset loading and preprocessing with PASCAL VOC 2007/2012 protocols.
- Letterbox resizing that preserves aspect ratio and adjusts ground truth boxes accordingly.
- Darknet-19 backbone with bottleneck and passthrough layers for multi-scale features.
- YOLOv2 detection head with decoding utilities to map raw logits into meaningful bounding boxes.
- Custom YOLOv2 loss, faithful to the original (coordinate, confidence, class terms).
- Postprocessing utilities including Non-Maximum Suppression (NMS).
- Training loop with gradient accumulation, mixed precision (AMP), and checkpointing.
- Evaluation with VOC-style mAP@0.5 (VOC07 11-point metric and continuous AP).
- Visualization tools to overlay bounding boxes, class labels, and confidence scores on images.
- A suite of unit tests ensuring correctness, reproducibility, and reliability.
The project is designed for clarity, modularity, and research reproducibility. It is a solid reference for understanding the YOLOv2 architecture, training it on VOC, or extending it to new datasets.
yolov2-implementation/
│
├─ src/
│ ├─ data/ # Datasets, transforms, encoding
│ ├─ models/ # Darknet-19, YOLOv2 body, layers
│ ├─ heads/ # YOLOv2 head, decoding, postprocess
│ ├─ losses/ # YOLOv2 loss function
│ ├─ inference/ # Forward inference + NMS
│ ├─ metrics/ # VOC mAP evaluation
│ ├─ train/ # Training loops and utilities
│ └─ tests/ # Unit tests (pytest)
│
├─ scripts/ # User-facing scripts (train, eval, visualize, demo)
│
├─ Notebooks/ # A full implementation in one jupytern notebook
│
├─ requirements.txt
├─ README.md
├─ LICENSE
YOLOv2 builds on the Darknet-19 backbone and introduces:
- Darknet-19: 19 convolutional layers with batch normalization and leaky ReLU.
- Passthrough layer: connects higher-resolution features (26x26) to the detection layer via space-to-depth transform.
- Detection head: predicts bounding boxes, objectness, and class probabilities directly from feature maps.
- Anchor boxes: predefined priors that allow YOLOv2 to generalize to different object shapes.
Input: 416x416 RGB image
Output: (B, A*(5+C), 13, 13) tensor, where A=anchors, C=classes
The repository includes scripts to train YOLOv2 on the PASCAL VOC datasets (2007+2012 for train/val, 2007 test).
Check the train.py file to see how the training loop is implemented, including support for gradient accumulation, mixed precision (AMP), and checkpoint saving/resume.
An evaluation pipeline is provided to compute mAP@0.5 on VOC datasets.
See the eval.py file for details on how detections are collected, NMS is applied, and AP per class is reported.
Visualization utilities allow plotting detections with bounding boxes, class labels, and confidence scores.
Refer to visualize.py for usage examples and customization.
YOLOv2 was a landmark in real-time object detection:
- Achieved state-of-the-art accuracy in 2017 with real-time inference.
- Introduced anchor boxes to the YOLO family.
- Balanced speed and accuracy, enabling deployment on constrained hardware.
- Influenced subsequent models like YOLOv3, YOLOv4, and beyond.
- Still relevant for educational, lightweight, and baseline benchmarking purposes.
- Redmon, J. and Farhadi, A. (2017). YOLO9000: Better, Faster, Stronger. CVPR.
- Original Darknet implementation: https://pjreddie.com/darknet/yolo/
- PyTorch: https://pytorch.org
This project is released under the MIT License. You are free to use, modify, and distribute it for research and commercial purposes, provided that proper attribution is given...