Incremental-DETR: Incremental Few-Shot Object Detection via Self-Supervised Learning

Incremental few-shot object detection aims at detecting novel classes without forgetting knowledge of the base with only a few labeled training data from classes. Most related prior works are on incremental that rely availability abundant samples per class substantially limits scalability to real-world setting where can be scarce. In this paper, we propose Incremental-DETR does via fine-tuning and self-supervised learning DETR detector. To alleviate severe over-fitting data, first fine-tune class-specific components self-supervision additional proposals generated using Selective Search as pseudo labels. We further introduce an strategy distillation encourage network in Extensive experiments conducted standard settings show our approach significantly outperforms state-of-the-art methods by large margin. Our source code is available https://github.com/dongnana777/Incremental-DETR.