Frame-Event Occlusion Tracking Dataset

v2.0 · 2025-05 Public · CC BY NC 4.0

Sequences

370 videos

Images

73K+ frames

Occlusions

11 levels

Challenges

10 attributes

Volume

93.4 GB

This is a high spatial resolution Frame-Event tracking dataset with pixel-level alignment and occlusion-level annotations, serving as a reliable benchmark for evaluating occlusion-robust single-object trackers. Feel free to download and use！
CVPR 2026 paper: Tracking through Severe Occlusion via Event-Derived Transient Cues.

Download FEOT Dataset

FEOT root folder
- Sequence_000001
  - event
    - 000001.png
    - 000002.png
    - ⋯
  - frame
    - 000001.png
    - 000002.png
    - ⋯
  - aligned_event.h5
  - gt.txt format per line: category_ID instance_ID left_x top_y width height occlusion_level
  - attribute.txt attribute IDs: OE(1), LI(2), RN(3), BI(4), VC(5), OC(6), FM(7), NM(8), SV(9), OV(10)
  - timestamp.txt unit: microseconds (μs)
- Sequence_000002
- Sequence_000003
- ⋯

Click to Download

Dataset Files

event folder

Event Frame Sequence

This folder contains a sequence of temporally continuous event frames generated from the event streams.

frame folder

RGB Frame Sequence

This folder contains a sequence of temporally ordered conventional images captured by an RGB camera.

aligned_event.h5

Aligned Event Stream

This file stores the event stream data in HDF5 format. The events have been spatially and temporally aligned with the frame images.

gt.txt

Ground Truth

This file contains frame-by-frame ground-truth bounding boxes of the target, annotated by a professional labeling company.

attribute.txt

Challenging Attributes

This file contains video attribute annotations: the first row lists attribute IDs, and the second row indicates presence (1) or absence (0).

timestamp.txt

Timestamps

This file contains the timestamps of each RGB image and event frame. The two modalities are temporally aligned.

Sample Preview

Supplementary Notes

Note on Dataset Sequence Count
The total number of sequences released in the current version (370 sequences) is slightly higher than the number reported in the original paper. This is primarily due to the inclusion of additional sequences that were rigorously filtered and added during the final stage of data curation and quality verification, aiming to enhance scene diversity and cover a broader range of occlusion patterns.
Privacy Protection Measures
To strictly safeguard any potential personal privacy, we have applied mosaic blurring to all regions containing human faces in the dataset. This processing step does not compromise the validity of the occlusion tracking annotations, while ensuring compliance with data privacy and ethical standards. The results reported in the paper are based on data before mosaicking.

Citation

@InProceedings{Dong2026CVPR,
        title = {Tracking through Severe Occlusion via Event-Derived Transient Cues},
        author = {Hao Dong, Yujin Liu, Haoyue Liu, Zhenyu Wang, Shihan Peng, Zhiwei Shi, Yi Chang, Luxin Yan},
        booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
        pages={XXX--XXX},
        year = {2026}
}

Statistics

Total Page Views

Avg. Stay Time

Paper Clicks

Dataset Clicks

Creative Commons Attribution 4.0 International (CC BY NC 4.0)

Maintainer: Hao Dong Contact Email: donghao0205@hust.edu.cn
Maintainer: Yujin Liu Contact Email: yujinliu@hust.edu.cn