Few-Shot Object Detection via Decoupled and Balanced Contrastive Learning

Pengxin Kang, Yuyong Cui, Xiaohe Cao, Dong Li, Zhenbao Luo, Huhai Jiang, Qi Zhang, Yi Shi, Zhonghe Tang
Proceedings of 2025 2nd International Conference on Machine Learning and Intelligent Computing, PMLR 278:305-312, 2025.

Abstract

Abstract. Only a few training examples are used for object detection task, and The productivity of the neural network model will show a dramatic decrease. Countless approaches to few-shot object detection (FSOD) have been formulated to solve the problem with a fine-tuning mechanism. Whereas those methods usually result in misclassification of novel classes and are biased in favor of base classes. For coping with this problem, we advance a fine-tuning learning framework with decoupled and balanced contrastive schemes(FSDB). More precisely, we first incorporate supervised contrastive learning with a decoupled loss to obtain a more outstanding performance for novel classes. Based on the decoupled supervised contrastive learning, we then put forward a class-balanced learning technique to resolve the issue of unequal sample distribution of base and new classes in the fine-tuning procedure. Rigorous experiments conducted on PASCAL VOC and MS-COCO datasets indicate that the presented technique has obtained excellent results for FSOD tasks.

Cite this Paper

BibTeX
@InProceedings{pmlr-v278-kang25a, title = {Few-Shot Object Detection via Decoupled and Balanced Contrastive Learning}, author = {Kang, Pengxin and Cui, Yuyong and Cao, Xiaohe and Li, Dong and Luo, Zhenbao and Jiang, Huhai and Zhang, Qi and Shi, Yi and Tang, Zhonghe}, booktitle = {Proceedings of 2025 2nd International Conference on Machine Learning and Intelligent Computing}, pages = {305--312}, year = {2025}, editor = {Zeng, Nianyin and Pachori, Ram Bilas and Wang, Dongshu}, volume = {278}, series = {Proceedings of Machine Learning Research}, month = {25--27 Apr}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v278/main/assets/kang25a/kang25a.pdf}, url = {https://proceedings.mlr.press/v278/kang25a.html}, abstract = {Abstract. Only a few training examples are used for object detection task, and The productivity of the neural network model will show a dramatic decrease. Countless approaches to few-shot object detection (FSOD) have been formulated to solve the problem with a fine-tuning mechanism. Whereas those methods usually result in misclassification of novel classes and are biased in favor of base classes. For coping with this problem, we advance a fine-tuning learning framework with decoupled and balanced contrastive schemes(FSDB). More precisely, we first incorporate supervised contrastive learning with a decoupled loss to obtain a more outstanding performance for novel classes. Based on the decoupled supervised contrastive learning, we then put forward a class-balanced learning technique to resolve the issue of unequal sample distribution of base and new classes in the fine-tuning procedure. Rigorous experiments conducted on PASCAL VOC and MS-COCO datasets indicate that the presented technique has obtained excellent results for FSOD tasks.} }
Endnote
%0 Conference Paper %T Few-Shot Object Detection via Decoupled and Balanced Contrastive Learning %A Pengxin Kang %A Yuyong Cui %A Xiaohe Cao %A Dong Li %A Zhenbao Luo %A Huhai Jiang %A Qi Zhang %A Yi Shi %A Zhonghe Tang %B Proceedings of 2025 2nd International Conference on Machine Learning and Intelligent Computing %C Proceedings of Machine Learning Research %D 2025 %E Nianyin Zeng %E Ram Bilas Pachori %E Dongshu Wang %F pmlr-v278-kang25a %I PMLR %P 305--312 %U https://proceedings.mlr.press/v278/kang25a.html %V 278 %X Abstract. Only a few training examples are used for object detection task, and The productivity of the neural network model will show a dramatic decrease. Countless approaches to few-shot object detection (FSOD) have been formulated to solve the problem with a fine-tuning mechanism. Whereas those methods usually result in misclassification of novel classes and are biased in favor of base classes. For coping with this problem, we advance a fine-tuning learning framework with decoupled and balanced contrastive schemes(FSDB). More precisely, we first incorporate supervised contrastive learning with a decoupled loss to obtain a more outstanding performance for novel classes. Based on the decoupled supervised contrastive learning, we then put forward a class-balanced learning technique to resolve the issue of unequal sample distribution of base and new classes in the fine-tuning procedure. Rigorous experiments conducted on PASCAL VOC and MS-COCO datasets indicate that the presented technique has obtained excellent results for FSOD tasks.
APA
Kang, P., Cui, Y., Cao, X., Li, D., Luo, Z., Jiang, H., Zhang, Q., Shi, Y. & Tang, Z.. (2025). Few-Shot Object Detection via Decoupled and Balanced Contrastive Learning. Proceedings of 2025 2nd International Conference on Machine Learning and Intelligent Computing, in Proceedings of Machine Learning Research 278:305-312 Available from https://proceedings.mlr.press/v278/kang25a.html.

Related Material