Date

2022-7

Author

Baltacı, Zeynep Sonat

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

547
views

136
downloads

Cite This

Objects are distributed unevenly in real world, which manifests itself as a long-tailed distribution in realistic visual recognition datasets. Deep learning based approaches trained on such imbalanced datasets using conventional gradient-based training strategies exhibit unfair recognition performances towards classes that are under-represented in the dataset. This so-called class imbalance has been studied in the literature by measuring imbalance via either class frequency or class hardness, and using those measures to mitigate imbalance by sampling, loss weighting or calibration strategies. In this thesis, we argue and empirically show that sample frequency or hardness alone is not sufficient for capturing imbalance among classes. Then we propose a novel measure based on predictive uncertainty of a trained deep network and demonstrate that it can capture imbalance better than existing approaches. Finally, we incorporate our measure to existing imbalance mitigation methods: loss reweighting, resampling, margin-based methods, and two-stage training. We show that predictive uncertainty-based methods improve over or perform on par with existing baselines on long-tailed datasets CIFAR-10-LT, CIFAR-100-LT and ImageNet-LT.

Subject Keywords

Long-tailed visual recognition, class imbalance, predictive uncertainty, imbalance mitigation

URI

https://hdl.handle.net/11511/98144

Collections

Graduate School of Natural and Applied Sciences, Thesis