Faster R-CNN Implementation with Pytorch

• Unofficial implementation of the paper Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks

0. Develop Environment

Docker Image
- pytorch/pytorch:1.8.1-cuda11.1-cudnn8-devel

• Using Single GPU

1. Implementation Details

• augmentation.py : augmentation considering images and bounding boxes
• dataset.py : VOC dataset using torchvision
• fast_rcnn.py : Fast R-CNN implementation
• faster_rcnn.py : Faster R-CNN implementation
• main.py : train, test, inference of Faster R-CNN
• rpn.py : Region Proposal Network implementation
• utils.py : class includes decode/encode bounding boxes coordinates, balancing samples, match predicts with ground truth
• train and inference.ipynb : install library, download dataset, preprocessing, train and result
• sample.png : sample image for visualize

2. Result Comparison on PASCAL VOC 2007

Source	Score	Detail
Paper	69.9	VOC 2007 trainval
Paper	73.2	VOC 2007 trainval + VOC 2012 trainval
Paper	78.8	VOC 2007 trainval + VOC 2012 trainval + COCO
Current Repo	67.5454	VOC 2007 trainval

3. Reference

• Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks [paper] [MATLAB official code] [Python official code]