Improving Noisy Fine-Grained Datasets using Active Label Cleaning Framework

Avik Pal; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 7273-7282

In this work we address the escalating data labeling challenge in deep learning focusing on the effectiveness of the Active Label Cleaning (ALC) framework in Fine-grained Visual Categorization (FGVC) tasks. With the rising complexity of models crowdsourcing becomes crucial but it often introduces noise. ALC leveraging Active Learning proves to be a cost-effective solution for relabeling specifically in FGVC datasets. The study explores acquisition functions for efficient sample prioritization and evaluates ALC's suitability in cleaning noisy FGVC data. Contributions made in this paper include simulating crowd-generated labels demonstrating ALC's efficacy in FGVC scenarios and highlighting its synergy with noise-robust learning methods. Prioritizing samples based on model posteriors and entropy emerges as a promising acquisition strategy.