Faster-RCNN Pytorch Implementaton

This is a simple implementation of Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. I mainly referred to two repositories below.

• longcw/faster_rcnn_pytorch
• jwyang/faster-rcnn.pytorch

Prerequisites

• python 3.5.x
• pytorch 0.4.1
• tensorboardX
• opencv3
• pillow
• easydict
• matplotlib

Purpose

• Training and testing on VOC
• ResNet-50
• ResNet-101

Main Results

voc07+12trainval/voc07test

original/res101	this/res101	this/res50
76.4	76.9	76.0

Running time: ~100ms(10FPS) on a GTX 1080

Visualization

You can inspect the accuray of both foreground and background samples in RPN and RCNN

Detection Results

Preparation

1. First clone the code

    git clone https://github.com/tztztztztz/faster-rcnn.pytorch.git
   

2. Install dependencies

    cd $PROJECT
    pip install -r requirements.txt
   

3. Compile roi_pooling layer and gpu_nms

    cd $PROJECT/faster_rcnn
    sh make.sh
   

Training on PASCAL VOC

Prepare the data

Please follow the instructions of this repository to prepare the data

Download the pretrained model on ImageNet

Get the model ResNet-50 ResNet-101, and put it at $PROJECT/data/pretrained folder. See more detail at ruotian/pytorch-resent

Train

python train.py --cuda

Evaluation

python test.py --cuda

you can check the detection results with command below

python test.py --cuda --vis