Attention Guided Cosine Margin For Overcoming Class-Imbalance in Few-Shot Road Object Detection

Few-shot object detection (FSOD) localizes and classifies objects in an image\ngiven only a few data samples. Recent trends in FSOD research show the adoption\nof metric and meta-learning techniques, which are prone to catastrophic\nforgetting and class confusion. To overcome these pitfalls in metric learning\nbased FSOD techniques, we introduce Attention Guided Cosine Margin (AGCM) that\nfacilitates the creation of tighter and well separated class-specific feature\nclusters in the classification head of the object detector. Our novel Attentive\nProposal Fusion (APF) module minimizes catastrophic forgetting by reducing the\nintra-class variance among co-occurring classes. At the same time, the proposed\nCosine Margin Cross-Entropy loss increases the angular margin between confusing\nclasses to overcome the challenge of class confusion between already learned\n(base) and newly added (novel) classes. We conduct our experiments on the\nchallenging India Driving Dataset (IDD), which presents a real-world\nclass-imbalanced setting alongside popular FSOD benchmark PASCAL-VOC. Our\nmethod outperforms State-of-the-Art (SoTA) approaches by up to 6.4 mAP points\non the IDD-OS and up to 2.0 mAP points on the IDD-10 splits for the 10-shot\nsetting. On the PASCAL-VOC dataset, we outperform existing SoTA approaches by\nup to 4.9 mAP points.\n