A framework for hyperspectral semantic segmentation based on pytorch and pytorch-lightning. This repository is part of a publication at 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2024) with the title

"HS3-Bench: A Benchmark and Strong Baseline for Hyperspectral Semantic Segmentation in Driving Scenarios"

Find the accepted version on arXiv: paper or on IEEEXplore paper

Supported Datasets

• HyKo2 (paper)
• HSI-Drive v2.0 (paper)
• Hyperspectral City V2.0
• HSI-Road (paper)
• WHU-OHS (paper)

Supported Models

• U-Net (paper)
• DeeplabV3+ (paper)
• RU-Net (regularized U-Net, ours)

Setup

Installation

• install Nvidia driver + CUDA (tested: driver 470.82/525.125.06 with CUDA 11.4/12.0)
• create environment using python=3.11
• pip install -e .

Dataset Preparation

1. Hyperseg: python scripts/prep_hyko.py

2. HSI-Drive: Can be downloaded here. For uncompressing the zip-file a password is required. Follow the instructions on the website

3. Hyperspectral City v2.0: Can be downloaded here. To use the our DataLoader it is required to decompress the data and convert it to a hdf5-file (see convert_hsds_to_hdf).

4. HSI-Road: Follow instructions on project page

5. WHUOHS: See project page or zenodo page.

Usage

Configure dataset basepaths in dataset configuration files (hyperseg/run/conf/dataset/<dataset>.yaml) and logdir (hyperseg/run/conf/<train/test_conf>.yaml). We use hydra for config management. You can customize train and test runs by overwriting parameters (field=value) when running the script in terminal. Hydras multi-run feature allows you to automatically run multiple experiments with different parameterizations (see conf-files and hydra docs for more information).

Training/Testing

Per default a vanilla unet and HyKo2 dataset is used. Models as well as datasplits can be found here.

# train
python run/train.py logging.project_name=<project_name>

# test
python run/test.py logging.project_name=<project_name> model.ckpt=<path_to_checkpoint>

Benchmark

Dataset	Approach	Data	$Acc_\mu$	$Acc_M$	$F_{1_{M}}$	$J_M$
HCV2	U-Net	HSI	85.25	48.62	48.18	37.73
	RU-Net	HSI	87.63	54.14	53.26	43.33
	RU-Net	PCA1	88.25	58.07	55.43	44.26
	RU-Net	pRGB	87.95	56.65	55.46	44.03
	DL3+	HSI	86.60	53.15	51.83	40.79
	DL3+	PCA1	86.64	54.46	52.90	41.58
	DL3+	pRGB	87.00	55.33	54.08	42.58
	DL3+(BB)	pRGB	90.26	64.10	61.93	50.04
	(DL3+(BB)	RGB	91.22	65.87	63.33	52.11)
	DL3+(PT)	pRGB	89.62	61.91	60.17	48.47
HyKo2	U-Net	HSI	85.36	68.15	68.55	57.39
	RU-Net	HSI	86.72	68.79	69.19	58.64
	RU-Net	PCA1	85.61	68.09	70.01	58.67
	RU-Net	pRGB	89.18	73.92	75.04	64.67
	DL3+	HSI	84.10	63.01	64.90	53.22
	DL3+	PCA1	79.99	61.59	63.00	50.40
	DL3+	pRGB	84.64	65.30	66.56	54.82
	DL3+(BB)	pRGB	90.49	74.87	77.11	66.77
	DL3+(PT)	pRGB	88.62	73.97	76.79	65.41
HSI-Drive	U-Net	HSI	94.95	74.74	76.08	64.95
	RU-Net	HSI	96.08	79.82	82.34	72.18
	RU-Net	PCA1	97.02	86.80	87.76	79.23
	RU-Net	pRGB	96.32	82.70	84.91	75.31
	DL3+	HSI	92.51	65.58	67.86	56.63
	DL3+	PCA1	90.88	62.93	64.31	52.62
	DL3+	pRGB	92.74	66.59	69.46	57.84
	DL3+(BB)	pRGB	97.09	83.93	86.41	77.44
	DL3+(PT)	pRGB	95.69	81.95	84.09	73.84
Average Performance	U-Net	HSI	88.52	63.84	64.27	53.36
	RU-Net	HSI	90.14	67.58	68.26	57.68
	RU-Net	PCA1	90.29	70.99	71.07	60.72
	RU-Net	pRGB	91.15	71.09	71.80	61.34
	DL3+	HSI	87.74	60.58	61.53	50.21
	DL3+	PCA1	85.84	59.66	60.07	48.20
	DL3+	pRGB	88.13	62.41	63.37	51.75
	DL3+(BB)	pRGB	92.61	74.30	75.15	64.75
	DL3+(PT)	pRGB	91.31	72.61	73.68	62.57
Worst-Case Performance	U-Net	HSI	82.25	48.63	48.18	37.73
	RU-Net	HSI	86.72	54.14	53.26	42.23
	RU-Net	PCA1	85.61	58.07	55.43	44.26
	RU-Net	pRGB	87.95	56.65	55.46	44.03
	DL3+	HSI	84.10	53.15	51.83	40.79
	DL3+	PCA1	79.99	54.46	52.90	41.58
	DL3+	pRGB	84.64	55.33	54.08	42.58
	DL3+(BB)	pRGB	90.26	64.10	61.93	50.04
	DL3+(PT)	pRGB	88.62	61.91	60.17	48.47

The results highlighted in italics where achieved by using additional data from an RGB cam and are therefore surrounded in parenthesis. DL3+ stands for DeepLabV3+. BB suffix means that we used a MobileNetV2 backbone network pretrained on ImageNet and then finetuned on pseudo-RGB images derived from HSI. PT suffix means that we used a MobileNetV2 backbone network pretrained on CityScapes without further fine-tuning. all layers except for output-layer were frozen.

Reference

If you use our code please reference our paper.

@INPROCEEDINGS{10801768,
  author={Theisen, Nick and Bartsch, Robin and Paulus, Dietrich and Neubert, Peer},
  booktitle={2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}, 
  title={HS3-Bench: A Benchmark and Strong Baseline for Hyperspectral Semantic Segmentation in Driving Scenarios}, 
  year={2024},
  volume={},
  number={},
  pages={5895-5901},
  keywords={Measurement;Learning systems;Protocols;Codes;Semantic segmentation;Training data;Benchmark testing;Standards;Intelligent robots;Hyperspectral imaging},
  doi={10.1109/IROS58592.2024.10801768}}