Faster RCNN Cpp

An implementation of Faster R-CNN with VGG16 backbone (https://arxiv.org/pdf/1506.01497.pdf) in libtorch.

Features

• Supports PASCAL VOC 2007 and MS COCO 2017 datasets
• Supports VGG16 backbone (from official torchvision model)
• Supports ROI Pool and ROI Align modes
• Supports PASCAL VOC 2007 and PASCAL VOC 2012 evaluation

Requirements

• libtorch
• torchvision
• OpenCV
• Boost

Build

cd faster-rcnn-cpp
cmake -S . -B build -DCMAKE_PREFIX_PATH=path_to_libtorch -DCMAKE_PREFIX_PATH=path_to_torchvision -DCMAKE_PREFIX_PATH=path_to_opencv -DCMAKE_PREFIX_PATH=path_to_boost
cmake --build build

Benchmark

PASCAL VOC 2007 (Train/Test: 07trainval/07test, non-difficult)

batch size: 1, lr: 0.00125, decay epoch: 9 12, total epoch: 12

ResNet50_FPN	AP	AP50
this	0.439	0.767
mmdet	0.437	0.769
ResNet50_FPN	0.438	0.768

VGG16	mAP
this	0.6798
Original Paper	0.699
VGG16 (Pytorch)	0.701

Todo

• Implement the dataset transform with the interface of libtorch
• Integrate evaluation into dataset and automatically run evaluation after test

Acknowledgement

This project is based on the following project:

• https://github.com/thisisi3/libtorch-faster-rcnn