Thermal Faster R-CNN for FIR Pedestrians Detection on SCUT

Introduction

This code is provided a modified Faster R-CNN for FIR pedestrians detection on SCUT Dataset.

The code RPN in this repo is written based on the MATLAB implementation of RPN+BF. Details about RPN+BF in: zhangliliang/RPN_BF.

The code external/code3.2.1 and external/toolbox is clone from SCUT-CV/SCUT_FIR_Pedestrian_Dataset which is based on Caltech dataset tool and Piotr's Image & Video Matlab Toolbox. Detials about Piotr’s Toolbox are in：pdollar/toolbox.

The SCUT FIR Pedestrians Dataset is a large far infrared pedestrian detection dataset. Detials about SCUT dataset are in: SCUT-CV.

Requirements

1. ubuntu (16.04 64bit)
2. MATLAB (our is MATLAB 2016b)
3. GPU: 1080ti or better

Installation

1. Clone the TFRCN reposityory

   git clone --recursive https://github.com/xzhewei/TFRCN.git

2. Build Caffe

   In ./external/caffe directory, there is our used caffe version. Follow the instruction to set up the prerequisites for Caffe. Use make matcaffe Build the mex file.

3. Download the SCUT Dataset

   • Download the SCUT Dataset the videos into ./external/code3.2.1/data-scut/videos directory
   • Download the SCUT Dataset the annotations into ./external/code3.2.1/data-scut/annotations directory

4. Download the VGG-16 pretrain model in VGG16_pretrain.zip from BaiduYun or GoogleDrive, and unzip it in the repo folder.

5. RUN ./startup() and ./tfrcn_build()

   ​

Training on SCUT

1. Start MATLAB from the repo folder

2. Training data preparation

   Extract image and annotation file into ./datasets

   extract_img_anno_scut('./external/code3.2.1/data-scut','./datasets/scut/')
   

3. Run script_tfrcn_train_scut to train and test the TFRCN model on SCUT. The result will auto copy into ./external/code3.2.1/data-scut/res

4. Run dbEval_scut , it would give the evaluation results on SCUT. The Reasonable MR is ~10%, Overall MR is ~33%.