Noise-Robust tiny object localization with flows

2026-01-01

Pattern Recognition

AI Summary

This work addresses the challenge that tiny object detection is highly susceptible to annotation noise, where stringent localization requirements often lead to overfitting. To mitigate this issue, the authors propose the TOLF framework, which introduces normalizing flows into tiny object localization for the first time to model the non-Gaussian distribution of prediction errors. Furthermore, an uncertainty-guided gradient modulation mechanism is designed to suppress learning from noisy samples. Evaluated on three benchmarks including AI-TOD, the proposed method significantly enhances detection performance, improving the AP of the DINO baseline by 1.2% on AI-TOD and demonstrating robustness in tiny object detection.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Object Detection & Categorization
• Search and Optimization: Learning to Search

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs

Problem

• tiny object detection
• annotation noise
• localization robustness
• noise overfitting
• object detection

Innovation

• normalizing flows
• tiny object detection
• annotation noise
• uncertainty-aware learning
• gradient modulation