Instance-level feature representation calibration for visual object detection

Hua Zhang, Jingzhi Li, Wenqi Ren, Chaopeng Li, Xiaochun Cao

Abstract: Highlights•The Highlights of this paper are summarized as follows: Prototype-Based Supervised Contrastive Learning: We introduced a novel prototype-based supervised contrastive learning method to reduce representation bias, enhancing the discriminative power of features for more accurate few-shot object detection.•Balanced Cross-Entropy Loss Function: We developed a balanced cross-entropy loss function that ensures stable detector performance, even with unbalanced sample sizes and limited data for novel classes, overcoming challenges in few-shot learning.•Experimental Validation on Benchmark Datasets: Extensive experiments on PASCAL VOC and MS-COCO benchmarks demonstrate the superior performance and effectiveness of our method compared to existing approaches.