RGBT-Tiny: A Large-Scale Benchmark for Visible-Thermal Tiny Object Detection

RGBT-Tiny is a large-scale visible-thermal benchmark, which consists of 115 high-quality paired image sequence, 93K frames and 1.2M manual annotations, and covers abundant targets and diverse senarios. Details of this dataset can be found in our paper. Over 81\% of targets are smaller than 16x16, and we provide paired bounding box annotations with tracking id to offer an extremely challenging benchmark with wide-range applications, such as RGBT fusion, detection and tracking.

Sample Videos

               

## Benchmark Properties ### Rich Diversity Fig. 1 (a) Target distribution in visible and thermal modalities. (b) Scene distribution (inner circle) across different light visions (outer circle).
### Large Density Variation Fig. 2 Density of each sequence. (x,y,z) are the numbers of sequences w.r.t. density levels (i.e., sparse, medium, dense).
### Small-Scale Targets Fig. 3 Size distribution of each target category.
### Temporal Occlusion Fig. 4 Temporal occlusion (i.e., no occlusion, slight occlusion, moderate occlusion, heavy occlusion).

## Evaluation Metric Fig. 5 An illustration of SAFit measure. (a) Pixels deviation between the center points of GT bbox and predicted bbox. (b) IoU-Deviation curves w.r.t different sizes of bboxes. (c)-(d) SAFit-Deviation curves under different C values.
### SAFit for evaluation Fig. 6 Comparisons among different measures for performance evaluation in visible and thermal modalities.
### SAFit loss for training SAFit results achieved by ATSS equipped with different losses in visible and thermal modalities of RGBT-Tiny dataset.
SAFit and IoU results achieved by ATSS equipped with different losses in COCO dataset.

## Baseline Results Table 1 SAFit-based results of existing visible detection (V-D), visible SOD (V-SOD), thermal SOD (T-SOD), visible-thermal detection methods (VT-D) methods on RGBT-Tiny dataset.
Table 2 IoU-based results of existing visible detection (V-D), visible SOD (V-SOD), thermal SOD (T-SOD), visible-thermal detection methods (VT-D) methods on RGBT-Tiny dataset.

## Downloads To access RGBT-Tiny dataset, please fill the following form: [https://forms.gle/EeRooNEYzXXporQt9](https://forms.gle/EeRooNEYzXXporQt9) ## Citiation ``` @article{RGBT-Tiny, journal = {arXiv preprint arXiv:2406.14482}, author = {Xinyi Ying and Chao Xiao and Ruojing Li and Xu He and Boyang Li and Zhaoxu Li and Yingqian Wang and Mingyuan Hu and Qingyu Xu and Zaiping Lin and Miao Li and Shilin Zhou and Wei An and Weidong Sheng and Li Liu}, title = {Visible-Thermal Tiny Object Detection: A Benchmark Dataset and Baselines}, year = {2024}, month = {6}, } ``` ## Contact Please contact us at ***yingxinyi18@nudt.edu.cn*** for any questions.