Training Robust Object Detectors From Noisy Category Labels and Imprecise Bounding Boxes.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Object detection performance is hindered by noisy training data, including inaccurate category labels and imprecise bounding boxes from crowd-sourcing platforms.
• Degenerated data quality significantly impairs the performance of standard object detection models.

Purpose of the Study:

• To propose a novel method, Meta-Refine-Net (MRNet), for training object detectors robustly from noisy category labels and imprecise bounding boxes.
• To develop a model-agnostic approach capable of learning from imperfect object detection data with minimal clean examples.

Main Methods:

• MRNet adaptively assigns lower weights to proposals with incorrect labels to mitigate classification errors.
• The method dynamically generates more accurate bounding box annotations, leveraging imprecise boxes for positive regression impacts.
• Refinement of bounding box annotations is achieved by jointly learning category and localization information.

Main Results:

• MRNet effectively suppresses large loss values from incorrect labels and alleviates misleading information from imprecise bounding boxes.
• The approach demonstrates improved accuracy in approximating ground-truth bounding boxes.
• Experiments on PASCAL VOC 2012 and MS COCO 2017 datasets validate the effectiveness and efficiency of MRNet.

Conclusions:

• MRNet offers a robust solution for training object detectors on noisy datasets, significantly improving performance.
• The model-agnostic nature and low requirement for clean data make MRNet a practical and efficient tool for real-world applications.