SCUT-Seg Dataset

A new benchmark dataset for thermal image semantic segmentation

Description

The SCUT-Seg Datasets is a large semantic segmentation dataset for thermal image. The image sequences are mainly from SCUT FIR Pedestrian Dataset. To ensure the diversity of the data, we further collected some images captured in the summer. Finally, we picked out 2,010 images with large scene gap as our final label images, and divide the objects in the image into 10 categories, including background, road, person, rider, car, truck, fence, tree, bus and pole.

Download

Image sequences and annotations

• BaiduYun code: 1234

• GoogleDrive

Annotations

The public annotation tool for driving scenes provided by hitachi automotive and industry lab is used to annotate our dataset. The label details of each category are as follows:

Name	Ground Truth value	Color
background	0	[0, 0, 0]
road	7	[128, 64, 128]
person	24	[60, 20, 220]
rider	25	[0, 0, 255]
car	26	[142, 0, 0]
truck	27	[70, 0, 0]
fence	13	[153, 153, 190]
tree	21	[35, 142, 107]
bus	28	[100, 60, 0]
pole	17	[153, 153, 153]

Citing SCUT-Seg Dataset

Please consider citing our paper in your publications if you find SCUT-Seg Dataset helps your research:

@article{XIONG2021103628,
title = {MCNet: Multi-level Correction Network for thermal image semantic segmentation of nighttime driving scene},
journal = {Infrared Physics & Technology},
pages = {103628},
year = {2021},
issn = {1350-4495},
doi = {https://doi.org/10.1016/j.infrared.2020.103628},
author = {Haitao Xiong and Wenjie Cai and Qiong Liu},
}

MCNet

Background

MCNet is a network architecture proposed in our paper MCNet: Multi-level Correction Network for thermal image semantic segmentation of nighttime driving scene,

by Haitao Xiong, Wenjie Cai, Qiong Liu from South China University of Technology.

Prerequisites

• PyTorch 1.0
  • pip3 install torch torchvision
• Easydict
  • pip3 install easydict
• Apex
• tqdm
  • pip3 install tqdm

Training

1. modify the config.py according to your requirements

2. train a network:

   export NGPUS=8
   python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py
   

Inference

1. evaluate a trained network on the validation set:

   python eval.py
   

2. input arguments:

   usage: -e epoch_idx -d device_idx [--verbose ] 
   [--show_image] [--save_path Pred_Save_Path]
   

Comparisons with SOTAs

The following results are referred from our paper. Recently, we found that training with mmsegmentation can achieve higher performance on the SCUT-Seg dataset. Therefore, we recommend researchers to train the SCUT-Seg dataset with mmsegmentation.

Method	Backbone	mIoU(%)
U-Net	----	41.53
FCN-32s	ResNet-101	48.43
FCN-16s	ResNet-101	59.65
DFN	ResNet-101	64.55
BiseNet	ResNet-101	63.03
PSPNet	ResNet-101	67.04
PSANet	ResNet-101	67.32
DeepLabV3	ResNet-101	64.64
DeepLabV3+	ResNet-101	68.00
MCNet	ResNet-101	69.79

Download Model

MCNet-ResNet101 code:1234

Code Borrowed From

Our code mainly refers from torchseg.

Contact

Please contact me xhtroman@gmail.com if you get stuck anywhere.