Using a ranking-based loss for long-tailed visual recognition

Date

2024-4-16

Author

Gülmez, Baran

Metadata

Show full item record

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

Item Usage Stats

436
views

310
downloads

Cite This

Long-tailed visual recognition, where certain classes contain far fewer samples than others, poses a critical challenge in learning-based computer vision applications. As real-world visual recognition datasets generally exhibit long-tailed distributions, addressing the challenge of learning in such long-tailed datasets is essential for many applications. In this thesis, for long-tailed visual recognition, we explore and adapt the Average Precision (AP) Loss, which was originally proposed by Chen et al. for the task of object detection. We found that the standard AP Loss performs similarly to traditional loss functions like Cross Entropy Loss on dealing with uneven class distributions. By introducing two specific modifications to AP Loss, we significantly improved the model's accuracy in identifying rare classes and its overall performance across all classes. We conducted thorough experiments to compare these improved AP Loss versions with other top-performing loss functions in the literature. Our findings showed that our modified AP Loss versions provide on par with or better performance than state-of-the-art loss functions.

Subject Keywords

Long-tailed visual recognition, Ranking based loss functions

URI

https://hdl.handle.net/11511/109478

Collections

Graduate School of Natural and Applied Sciences, Thesis