MLC-NC: Long-Tailed Multi-Label Image Classification Through the Lens of Neural Collapse
Authors: Zijian Tao, Shao-Yuan Li, Wenhai Wan, Jinpeng Zheng, Jia-Yao Chen, Yuchen Li, Sheng-Jun Huang, Songcan Chen
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on widely-used benchmarks demonstrate the effectiveness of our method. Tables 1 and 2 present the experimental results on COCO-LT, VOC-LT, and VG200. Our MLC-NC demonstrates significant performance improvements over the second-best baseline, especially in the medium and tail classes, and achieves the best overall total performance. We first conduct ablation studies on the three key components of our method. |
| Researcher Affiliation | Academia | 1MIIT Key Laboratory of Pattern Analysis and Machine Intelligence 2College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics 3State Key Laboratory for Novel Software Technology, Nanjing University 4School of Computer Science and Technology, Huazhong University of Science and Technology 5College of Computer and Software, Hohai University |
| Pseudocode | Yes | The pseudocode for MLC-NC is provided in Appendix A. |
| Open Source Code | No | The paper does not contain an explicit statement about releasing code or a link to a code repository. |
| Open Datasets | Yes | Dataset: We analyze and conduct experiments on two artificially created long-tailed multi-label image classification datasets VOC-LT and COCO-LT following(Wu et al. 2020). Besides, we verify the universality of the proposed approach on one real-world multi-label dataset VG200 with milder imbalance distribution(Krishna et al. 2017). |
| Dataset Splits | No | The paper mentions using COCO-LT, VOC-LT, and VG200 datasets and discusses training instances and evaluation with m AP, but it does not specify explicit training, validation, or test split percentages or sample counts for these datasets. |
| Hardware Specification | No | The paper mentions using a ResNet50 pre-trained on ImageNet as a feature extractor, but it does not specify any particular hardware (GPU, CPU models, etc.) used for running their experiments. |
| Software Dependencies | No | The paper does not list any specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | Training Setup: We use a Res Net50 pre-trained on Image Net as the feature extractor. In LFPC, τ1 and τ2 are set to 0.5. α and β are set to 0.5 and 0.2, The dimension d of the feature projection is set to 20. We evaluate mean average precision (m AP) across all classes, averaging the results over three runs for all methods. |