Reproduce “Portable Active Learning for Object Detection” on OpenResearch

About this paper

Rashi Sharma, Justin Timothy C. Bersamin, Karthikk Subramanian

Portable Active Learning (PAL) is introduced as a detector-agnostic framework for object detection, selecting informative samples for annotation purely from inference outputs. It consistently achieved higher detection accuracy while requiring fewer labeled samples across multiple datasets and detector architectures compared to existing active learning methods, making it easier to integrate and deploy.

Key insights

• A detector-agnostic active learning framework can be highly effective and broadly compatible by relying exclusively on inference outputs.
• A synergistic combination of instance-level uncertainty with image-level diversity, global uncertainty, and class-imbalance awareness leads to more efficient and accurate data selection.
• Simple logistic classifiers prove more robust and effective for instance uncertainty estimation compared to complex models, particularly for low-frequency classes, preventing overfitting and ensuring reliable performance.

Summary by alphaXiv.