Visible-Thermal Tiny Object Detection: A Benchmark Dataset and Baselines.

Area of Science:

• Computer Vision
• Machine Learning
• Sensor Fusion

Background:

• Visible-thermal small object detection (RGBT SOD) is crucial for applications like surveillance and rescue but lacks dedicated benchmarks.
• Existing datasets are limited in scale, diversity, and object size, hindering impartial algorithm evaluation.
• Current RGBT SOD research is underdeveloped, with most studies focusing on single modalities.

Purpose of the Study:

• Introduce RGBT-Tiny, the first large-scale, diverse benchmark for RGBT SOD.
• Provide a challenging dataset with numerous small objects and varied scenes for robust algorithm testing.
• Propose a novel evaluation metric, Scale Adaptive Fitness (SAFit), for improved RGBT SOD performance assessment.

Main Methods:

• Constructed RGBT-Tiny dataset: 115 sequences, 93K frames, 1.2M annotations, 7 object categories, 8 scene types.
• Annotated over 81% of objects smaller than 16x16 pixels with bounding boxes and tracking IDs.
• Developed SAFit, a robust metric for evaluating RGBT SOD performance across small and large objects.

Main Results:

• RGBT-Tiny dataset enables comprehensive evaluation of RGBT SOD algorithms.
• SAFit metric demonstrates high robustness and promotes detection performance compared to IoU.
• Extensive evaluations of 30 state-of-the-art algorithms were conducted on the RGBT-Tiny benchmark.

Conclusions:

• RGBT-Tiny serves as a vital resource for advancing RGBT SOD research.
• The proposed SAFit metric enhances the reliability of RGBT SOD algorithm evaluation.
• This work establishes a new standard for benchmarking and developing RGBT SOD techniques.