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

Images will inevitably contain objects at extremely tiny scales when observations approach the camera's physical limit. Despite extreme, this scenario is ubiquitous in real world from micro-vision (medical, cell imaging) to macro-vision (drone, satellite imagery). In these specialized domains, imaging typically adopts a bird-eye's view to capture object's primary features, making them show arbitrary orientations. These oriented tiny objects pose severe challenges to object detection:

• Difficult Label Acquisition: The limited appearance information makes the labelling quite difficult.
• Weak Feature Representation: The features of oriented tiny objects will be lost in deep networks.
• Dense Object Arrangement: There can be thousands of objects per image due to overhead view.

These challenges give rise to the following problems:

• Existing methods struggle to deliver satisfactory performance. 77% objects in DOTA-v2 are in the size range of $10^{2}$-$50^{2}$ pixels, while the SOTA performance is still lower than 30% $\rm{AP^{@50:5:95}}$.
• Detecting oriented tiny objects is a hard nut to crack in the community, while there is a lack of task-specific datasets and benchmarks to prompt the development of detection methods.

The objects' tiny size and low confidence make them easily suppressed or ignored during model training. The vanilla optimization process will inevitably pose them into significantly biased prior setting and biased sample learning dilemmas, severely impeding the detection performance:

• Biased prior setting: Most prior positions are deviated from tiny objects' main area under fixed prior setting constraint by feature stride (e.g., 8, 16, 32, 64).
• Biased sample learning: Oriented tiny objects are assigned with much fewer number of positive samples compared to otehr objects due to sub-optimal measaurement and assignment strategy.

Towards scale-unbiased sample learning, we reformulate the training process into a Dynamic Coarse-to-Fine Learning (DCFL) pipeline with:

• Dynamic prior setting: Prior positions are dynamically updated to better accommodate tiny objects' extreme shapes.
• Coarse-to-fine sample learning: The coarse step warrants sufficient and diverse positive samples for each object. The fine step guarantees the learning quality by fitting each gt with a Dynamic Gaussian Mixture Model as a constraint to select low-quality samples.