Oriented Tiny Object Detection: A Dataset, Benchmark, and Dynamic Unbiased Learning.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Detecting oriented tiny objects is challenging due to limited appearance information.
• Existing methods struggle with the prevalence of these objects in real-world applications.

Purpose of the Study:

• Introduce a new dataset (AI-TOD-R) and benchmark for oriented tiny object detection.
• Propose a Dynamic Coarse-to-Fine Learning (DCFL) scheme to mitigate learning biases.

Main Methods:

• Developed AI-TOD-R, a dataset featuring the smallest oriented objects.
• Created a benchmark covering supervised and label-efficient detection methods.
• Implemented DCFL to dynamically update object priors and balance sample selection.

Main Results:

• Identified a learning bias where confident objects become more confident, marginalizing tiny objects.
• DCFL effectively mitigates this bias by improving prior alignment and sample balancing.
• Achieved state-of-the-art accuracy, efficiency, and versatility across 10 datasets.

Conclusions:

• The proposed DCFL scheme significantly enhances oriented tiny object detection.
• AI-TOD-R and the benchmark provide valuable resources for advancing the field.
• The findings offer a pathway to more robust and unbiased object detection models.