Repository for PST900 RGB-Thermal Calibration, Dataset and Segmentation Network | C++, Python, PyTorch
If you have any questions regarding this dataset, please raise a GitHub issue here or reach out to us at [email protected] (Shreyas S. Shivakumar) or [email protected] (Neil Rodrigues).
If you use this dataset, please cite our arXiv paper below:
@misc{shivakumar2019pst900,
title={PST900: RGB-Thermal Calibration, Dataset and Segmentation Network},
author={Shreyas S. Shivakumar and Neil Rodrigues and Alex Zhou and Ian D. Miller and Vijay Kumar and Camillo J. Taylor},
year={2019},
eprint={1909.10980},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
If you wish to cite our overall DARPA Subterranean Challenge autonomy system, please cite:
@misc{miller2019tunnel,
title={Mine Tunnel Exploration using Multiple Quadrupedal Robots},
author={Ian D. Miller and Fernando Cladera and Anthony Cowley and Shreyas S. Shivakumar and Elijah S. Lee and Laura Jarin-Lipschitz and Akhilesh Bhat and Neil Rodrigues and Alex Zhou and Avraham Cohen and Adarsh Kulkarni and James Laney and Camillo Jose Taylor and Vijay Kumar},
year={2019},
eprint={1909.09662},
archivePrefix={arXiv},
primaryClass={cs.RO}
}
Dataset | Total Samples | Train/Test Split | Download Link | Size |
---|---|---|---|---|
PST900 RGB-T | 894 | Yes | Google Drive | 1.4GB |
PST-RGB | 3359* | No | Google Drive | 3.1GB |
* our paper mentions 3416 annotated RGB images. Additional images will be provided on request.
The directory structure for this dataset is as follows:
.
├── test/ # Train data split (606 pairs)
| ├── rgb/ # RGB Images
| | ├── 31_bag1a_rect_rgb_frame0000000007.png
| | ├── ...
| ├── thermal/ # Thermal Images (FLIR's AGC 8-bit)
| | ├── 31_bag1a_rect_rgb_frame0000000007.png
| | ├── ...
| ├── thermal_raw/ # Thermal Images (RAW 16-bit)
| | ├── 31_bag1a_rect_rgb_frame0000000007.png
| | ├── ...
| ├── depth/ # Depth image from stereo depth estimation
| | ├── 31_bag1a_rect_rgb_frame0000000007.png
| | ├── ...
| ├── labels/ # Human annotated per-pixel label
| | ├── 31_bag1a_rect_rgb_frame0000000007.png
| | ├── ...
├── train/ # Test data split (288 pairs)
| ├── rgb/ # RGB Images
| | ├── 31_bag1a_rect_rgb_frame0000000007.png
| | ├── ...
| ├── thermal/ # Thermal Images (FLIR's AGC 8-bit)
| | ├── 31_bag1a_rect_rgb_frame0000000007.png
| | ├── ...
| ├── thermal_raw/ # Thermal Images (RAW 16-bit)
| | ├── 31_bag1a_rect_rgb_frame0000000007.png
| | ├── ...
| ├── depth/ # Depth image from stereo depth estimation
| | ├── 31_bag1a_rect_rgb_frame0000000007.png
| | ├── ...
| ├── labels/ # Human annotated per-pixel label
| | ├── 31_bag1a_rect_rgb_frame0000000007.png
| | ├── ...
The directory structure for this dataset is as follows:
├── rgb/ # RGB Images
| ├── 01_levine_rgb_1_rdb_bag_100109
| ├── ...
├── labels/ # Human annotated per-pixel label
| ├── 01_levine_rgb_1_rdb_bag_100109
| ├── ...
We provide a basic set of tools for working with this dataset. This tools are mainly focused on how to read and represent the RGB, Thermal, Thermal (16-bit) and Labels.
Basic utilities can be found in the toolkit/utilities.py script.
A demo of the utilities are provided in utilities.py; this can be run as a python script. Provide the location to the dataset to the pst900_path variable and run the following command:
python3 utilities.py
Once you select a split type (either Train or Test, or another custom split), you can use our dataset sample fetch script:
# Example dataset sample loader
rgb, depth, thermal, thermal_raw, label = utils.get_sample(102)
and each modality should look like this:
==== Loading sample ====
Loaded RGB image | H: 720, W: 1280, C: 3, Type: uint8
Loaded Depth image | H: 720, W: 1280, C: 1, Type: uint16
Loaded Thermal image | H: 720, W: 1280, C: 1, Type: uint8
Loaded Thermal Raw image | H: 720, W: 1280, C: 1, Type: uint16
Loaded label image | H: 720, W: 1280, C: 1, Type: uint8
Note: Depth and Thermal Raw (16 bit) must be loaded as 16-bit images.
The RGB and Thermal (8-bit) can be easily viewed and should look like this:
The following method will perform a simple hole filling in the thermal images. The holes are a result of projecting the information from the Thermal camera onto the RGB camera. The holes are a result of parallax, resulting in a many-to-one mapping of pixels in RGB into the Thermal camera plane.
# Example hole filling for Thermal image
thermal_filled = utils.fill_thermal(thermal)
The output should look like this:
The labels are stored as uint8 images, with each pixel representing the class index {0..4}. This class indices can be mapped to colors for easier visualization as follows:
# Example colormapping for Label image
colormapped_label = utils.visualize_label(label)
The resulting image should look like this: