This repository provides a PyTorch implementation of the YOLOV7 object detection model, enabling users to train custom datasets. It offers comprehensive tools for training, prediction, and evaluation, making it suitable for researchers and developers working on object detection tasks.

How It Works

The implementation is based on the YOLOV7 architecture, known for its efficiency and accuracy. It supports various training configurations, including different learning rate schedulers (step, cosine), optimizers (Adam, SGD), and adaptive learning rate adjustments based on batch size. The code also incorporates features like image cropping, multi-GPU training, and class-specific object count calculations.

Quick Start & Requirements

• Install: pip install -r requirements.txt (specific torch version recommended)
• Prerequisites: PyTorch (version 1.7.1+ recommended for AMP), Python.
• Data: Requires VOC format datasets. Pre-trained weights and VOC dataset links are provided via Baidu Netdisk.
• Docs: Training, prediction, and evaluation steps are detailed in the README.

Highlighted Details

• Achieves mAP of 50.7 (0.5:0.95) and 69.2 (0.5) on COCO-Val2017 using yolov7_weights.pth.
• Supports yolov7_x variant with higher mAP (52.4 / 70.5).
• Includes adaptive learning rate adjustment based on batch size.
• Supports EMA (Exponential Moving Average) for improved training stability.

Maintenance & Community

The repository is maintained by bubbliiiing. Related repositories for YoloV3, YoloV4, YoloV5, and YoloX are also available from the same author.

Licensing & Compatibility

The repository does not explicitly state a license in the README. It is a PyTorch implementation, generally compatible with commercial use if the underlying YOLOV7 license permits.

Limitations & Caveats

The primary dependency is PyTorch version 1.2.0, though higher versions (1.7.1+) are recommended for AMP. Data download links are via Baidu Netdisk, which may pose accessibility issues for some users. The README mentions that directly resizing images without letterboxing might yield better results, suggesting potential tuning required for optimal performance.