Improving tail-class representation with centroid contrastive learning
In vision domain, large-scale natural datasets typically exhibit long-tailed distribution which has large class imbalance between head and tail classes. This distribution poses difficulty in learning good representations for tail classes. Recent developments have shown good long-tailed model can be...
| Main Authors: | , , , , , |
|---|---|
| Other Authors: | |
| Format: | Journal Article |
| Language: | English |
| Published: |
2023
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/172214 |
| _version_ | 1826113159485718528 |
|---|---|
| author | Tiong, Anthony Meng Huat Li, Junnan Lin, Guosheng Li, Boyang Xiong, Caiming Hoi, Steven C. H. |
| author2 | School of Computer Science and Engineering |
| author_facet | School of Computer Science and Engineering Tiong, Anthony Meng Huat Li, Junnan Lin, Guosheng Li, Boyang Xiong, Caiming Hoi, Steven C. H. |
| author_sort | Tiong, Anthony Meng Huat |
| collection | NTU |
| description | In vision domain, large-scale natural datasets typically exhibit long-tailed distribution which has large class imbalance between head and tail classes. This distribution poses difficulty in learning good representations for tail classes. Recent developments have shown good long-tailed model can be learnt by decoupling the training into representation learning and classifier balancing. However, these works pay insufficient consideration on the long-tailed effect on representation learning. In this work, we propose interpolative centroid contrastive learning (ICCL) to improve long-tailed representation learning. ICCL interpolates two images from a class-agnostic sampler and a class-aware sampler, and trains the model such that the representation of the interpolative image can be used to retrieve the centroids for both source classes. We demonstrate the effectiveness of our approach on multiple long-tailed image classification benchmarks. |
| first_indexed | 2024-10-01T03:18:37Z |
| format | Journal Article |
| id | ntu-10356/172214 |
| institution | Nanyang Technological University |
| language | English |
| last_indexed | 2024-10-01T03:18:37Z |
| publishDate | 2023 |
| record_format | dspace |
| spelling | ntu-10356/1722142023-11-29T06:15:24Z Improving tail-class representation with centroid contrastive learning Tiong, Anthony Meng Huat Li, Junnan Lin, Guosheng Li, Boyang Xiong, Caiming Hoi, Steven C. H. School of Computer Science and Engineering Engineering::Computer science and engineering Imbalanced Learning Contrastive Learning In vision domain, large-scale natural datasets typically exhibit long-tailed distribution which has large class imbalance between head and tail classes. This distribution poses difficulty in learning good representations for tail classes. Recent developments have shown good long-tailed model can be learnt by decoupling the training into representation learning and classifier balancing. However, these works pay insufficient consideration on the long-tailed effect on representation learning. In this work, we propose interpolative centroid contrastive learning (ICCL) to improve long-tailed representation learning. ICCL interpolates two images from a class-agnostic sampler and a class-aware sampler, and trains the model such that the representation of the interpolative image can be used to retrieve the centroids for both source classes. We demonstrate the effectiveness of our approach on multiple long-tailed image classification benchmarks. Nanyang Technological University National Research Foundation (NRF) Anthony Meng Huat Tiong is supported by Salesforce and the Singapore Economic Development Board under the Industrial Postgraduate Programme. Boyang Li is supported by the Nanyang Associate Professorship and the National Research Foundation Fellowship (NRF-NRFF13-2021-0006), Singapore. 2023-11-29T06:15:24Z 2023-11-29T06:15:24Z 2023 Journal Article Tiong, A. M. H., Li, J., Lin, G., Li, B., Xiong, C. & Hoi, S. C. H. (2023). Improving tail-class representation with centroid contrastive learning. Pattern Recognition Letters, 168, 123-130. https://dx.doi.org/10.1016/j.patrec.2023.03.010 0167-8655 https://hdl.handle.net/10356/172214 10.1016/j.patrec.2023.03.010 2-s2.0-85150821315 168 123 130 en NRF-NRFF13-2021-0006 Pattern Recognition Letters © 2023 Elsevier B.V. All rights reserved. |
| spellingShingle | Engineering::Computer science and engineering Imbalanced Learning Contrastive Learning Tiong, Anthony Meng Huat Li, Junnan Lin, Guosheng Li, Boyang Xiong, Caiming Hoi, Steven C. H. Improving tail-class representation with centroid contrastive learning |
| title | Improving tail-class representation with centroid contrastive learning |
| title_full | Improving tail-class representation with centroid contrastive learning |
| title_fullStr | Improving tail-class representation with centroid contrastive learning |
| title_full_unstemmed | Improving tail-class representation with centroid contrastive learning |
| title_short | Improving tail-class representation with centroid contrastive learning |
| title_sort | improving tail class representation with centroid contrastive learning |
| topic | Engineering::Computer science and engineering Imbalanced Learning Contrastive Learning |
| url | https://hdl.handle.net/10356/172214 |
| work_keys_str_mv | AT tionganthonymenghuat improvingtailclassrepresentationwithcentroidcontrastivelearning AT lijunnan improvingtailclassrepresentationwithcentroidcontrastivelearning AT linguosheng improvingtailclassrepresentationwithcentroidcontrastivelearning AT liboyang improvingtailclassrepresentationwithcentroidcontrastivelearning AT xiongcaiming improvingtailclassrepresentationwithcentroidcontrastivelearning AT hoistevench improvingtailclassrepresentationwithcentroidcontrastivelearning |