Detection of novel objects from a few annotated examples, known as few-shot object detection (FSOD), is highly desirable and received significant interest from researchers but remains challenging for modern systems.
Research has shown good feature embedding is key to good performance.
However, many systems still use Euclidean space, although hyperbolic space better encodes the data's hierarchical information.
Another way to optimize feature embeddings is contrastive learning which promotes intra-class similarity and inter-class difference.
We propose Hyperbolic and Contrastive Embedding using Decoupling with Baby Learning (HyCo-DeB) a novel few-shot object detector that realizes the unused potential in data for FSOD by optimizing the feature embeddings using hyperbolic space through a novel hyperbolic classification head and a contrastive head.
HyCo-DeB addresses the increased complexity of these heads and the transition in transfer-learning by using Baby Learning, allowing it to first transition to the new task, then gradually increase the complexity.
Being based on the widely used Faster R-CNN, our model deals with the conflicts of the class-agnostic RPN and the class-relevant RCNN head that shares the same backbone and the conflict in localization and classification by decoupling the modules.
Experiments show HyCo-DeB outperforms the existing state-of-the-art on the Pascal VOC benchmark and is on par with state-of-the-art on the MS COCO benchmark.