Unc-SOD: An Uncertainty Learning Framework for Small Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Small object detection (SOD) is challenging due to limited informative regions and inherent ambiguity in small instances.
• Existing two-stage detectors struggle with the uncertainty and feature inconsistency in SOD.
• Current sampling methods may be misled by unrecognizable small targets, wasting computational resources.

Purpose of the Study:

• To develop an Uncertainty learning framework for Small Object Detection (Unc-SOD).
• To address the challenges of uncertainty and feature inconsistency in small object detection.
• To improve the accuracy and efficiency of detecting small objects in complex scenes.

Main Methods:

• Incorporated an auxiliary uncertainty branch into the conventional Region Proposal Network (RPN) to model instance-level indeterminacy.
• Utilized uncertainty as a dynamic sampling criterion for proposal networks, improving candidate selection.
• Devised a Perception-and-Interaction strategy to capture rich and discriminative representations by optimizing regional features.

Main Results:

• Unc-SOD achieved state-of-the-art performance on the SODA-D and SODA-A benchmarks.
• The framework demonstrated significant improvements on COCO, VisDrone, and Tsinghua-Tencent 100K datasets.
• Results indicate superior performance compared to prevailing detectors when handling small instances.

Conclusions:

• Unc-SOD effectively models and leverages uncertainty for improved small object detection.
• The proposed framework offers a robust solution for the challenging SOD task.
• This work represents a seminal attempt to incorporate uncertainty modeling in SOD, yielding significant performance gains.