Related Experiment Video
Updated: Jul 27, 2025
06:25
Author Spotlight: Assessment of Visual Acuity in Central Vision Loss Through Motion-Based Peripheral Vision Testing
Published on: February 23, 2024
637
Divide and Retain: A Dual-Phase Modeling for Long-Tailed Visual Recognition.
Summary
This study addresses gradient distortion in visual recognition models trained on imbalanced long-tailed datasets. A novel dual-phase method disentangles gradients for data-rich and data-poor classes, improving model training and performance.
Area of Science:
- Computer Vision
- Machine Learning
- Deep Learning
Background:
- Real-world visual recognition datasets often exhibit long-tailed distributions, where a few classes have abundant data and many classes have scarce data.
- Existing methods typically train models holistically, leading to gradient distortion where gradients are biased towards data-rich classes and destabilized by data-poor classes.
- This gradient distortion problem hinders the effective training of visual recognition models on imbalanced datasets.
Purpose of the Study:
- To propose a novel method that mitigates gradient distortion in visual recognition models trained on long-tailed datasets.
- To improve the performance of visual recognition models by addressing the challenges posed by extreme data imbalance.
- To develop a training strategy that effectively handles both data-rich and data-poor classes.
Main Methods:
- A dual-phase training approach is introduced to disentangle gradients from data-rich and data-poor classes.
- Phase I focuses on updating model parameters using gradients from data-rich classes only.
- Phase II incorporates data-poor classes to learn a complete classifier, utilizing an exemplar bank and a memory-retentive loss for smooth transition and parameter stability.
Main Results:
- The proposed dual-phase method significantly reduces the negative impact of gradient distortion.
- Experimental results on CIFAR100-LT, Places-LT, ImageNet-LT, and iNaturalist 2018 datasets demonstrate superior performance compared to existing approaches.
- The exemplar bank and memory-retentive loss effectively maintain information from data-rich classes and stabilize parameter updates during the transition between phases.
Conclusions:
- The proposed dual-phase method offers an effective solution for training visual recognition models on long-tailed datasets.
- Disentangling gradients and employing strategic phase-based training with memory mechanisms improves model accuracy and robustness.
- This work provides a valuable contribution to the field of imbalanced learning in computer vision.
