// the find
bubbliiiing/faster-rcnn-pytorch
这是一个faster-rcnn的pytorch实现的库，可以利用voc数据集格式的数据进行训练。
A PyTorch implementation of Faster R-CNN for object detection, aimed at people learning the two-stage detection pipeline or who need a working baseline to train on VOC-format datasets. The code includes ResNet50 and VGG16 backbones with pretrained weights, plus prediction, training, and mAP evaluation scripts. It's a teaching repo first, a production tool second.
The code is broken into sensible modules (RPN, classifier, backbone, anchor utils) which makes it easier to follow than the original reference implementations. Supports both ResNet50 and VGG16 backbones out of the box with documented mAP numbers on VOC07+12 (80.36 and 77.46 respectively). Training is configurable with SGD/Adam, cosine/step LR decay, and batch-size-adaptive LR, which are non-trivial to wire up correctly. The VOC annotation tooling and data split logic is self-contained and straightforward to adapt for custom datasets.
Requires torch==1.2.0, which is roughly five major versions behind current — expect compatibility issues with modern CUDA drivers and any other library in your environment. Pretrained weights are hosted on Baidu Pan, which is inaccessible without a Chinese phone number for most people outside mainland China. The repo has been unmaintained since October 2023 and never addressed modern detection improvements like FPN or anchor-free heads, so the architecture is purely classical 2019-era Faster R-CNN. No COCO support either — you're stuck with VOC format, which limits how easily you can benchmark against anything published in the last few years.