Pol-CACTI is a monochrome polarized video dataset for research in polarized video snapshot compressive imaging. The accompanying paper presents an optical system that encodes polarization dynamics into a single measurement, together with a reconstruction network that jointly recovers temporal information and polarization channels. Its Polarization-Aware Learnable Back-Projection (PA-LBP) module promotes consistency with DoFP measurements. See the paper for the acquisition, Stokes-domain denoising, and reconstruction methodology.

The frame-based data described here contain 8,889 frames organized into 141 sequence folders, with separate training and testing splits.

Dataset Overview

Split	Sequence folders	Frames	Frames per folder	Image resolution (H × W)	File format
Train	131	8,409	30–92	2048 × 2048	BMP
Test	10	480	48	1024 × 1024	BMP
Total	141	8,889	—	—	—

Each folder contains one image sequence.
All image files store single-channel, 8-bit (uint8) arrays.
The current test split contains 480 BMP files .
Sequence lengths vary in the training split. Each test sequence contains exactly 48 frames.
Keep the provided Train and Test split assignments when reporting results.

The resolution listed above is the size of each stored image array. It does not denote the resolution of an individual polarization-angle channel.

Acquisition Hardware

The dataset was captured using a FLIR Blackfly S BFS-U3-51S5P-C polarization camera equipped with a Sony IMX250MZR division-of-focal-plane (DoFP) sensor.

Property	Specification
Camera	FLIR Blackfly S BFS-U3-51S5P-C
Sensor	Sony IMX250MZR
Imaging modality	Monochrome polarization
Shutter	Global shutter
Native camera resolution	2448 × 2048 pixels
Nominal full-resolution maximum frame rate	75 FPS (camera specification)
Linear polarization analyzer angles	0°, 45°, 90°, and 135°

Hardware specifications are available from the camera manufacturer. The nominal camera frame rate is a hardware specification; the acquisition rate and effective reconstruction frame rate should be taken from the experimental protocol.

Released-image DoFP angle layout: to be documented. The exact mapping of pixel positions to analyzer angles should be confirmed for the exported images before polarization demosaicing or Stokes computation.

Download

Dataset download: Hugging Face — Pol-CACTI

The data package contains the following two directories:

dataset/
├── Train/
│   ├── <sequence_id>/
│   │   ├── <frame_id>.bmp
│   │   └── ...
│   └── ...
└── Test/
    ├── <sequence_id>/
    │   ├── <frame_id>.bmp or <frame_id>.png
    │   └── ...
    └── ...

Use the original sequence and frame filenames. Filenames may contain a frame_ prefix, and frame indices may have gaps. Sort the available frames by their numeric index rather than assuming consecutive filenames.

Citation

If you use the dataset in your research, please cite:

@inproceedings{song2026polcacti,
  author    = {Song, Yunfeng and Luo, Yidong and Wang, Ping and
               Jiang, Xingjian and Yuan, Xin},
  title     = {{Pol-CACTI}: A System and Dataset for High-Speed
               Polarized Video Compressive Imaging},
  booktitle = {Computer Vision -- ECCV 2026},
  year      = {2026},
  volume    = {17028},
  series    = {Lecture Notes in Computer Science},
  pages     = {521--539},
  publisher = {Springer, Cham},
  doi       = {10.1007/978-3-032-37624-4_31},
  url       = {https://doi.org/10.1007/978-3-032-37624-4_31}
}

License

Dataset license: to be specified by the authors.

Contact

For questions about the dataset, please contact: songyunfeng@westlake.edu.cn

Downloads last month: 60

Total file size:

35.8 GB