MINI: Mining Implicit Novel Instances for Few-Shot Object Detection

Yuhang Cao, Jiaqi Wang, Yiqi Lin, Dahua Lin

Keywords: Object Detection, Few-Shot Object Detection

Abstract: Few-Shot Object Detection (FSOD) aims to detect novel concepts given abundant base data and limited novel data. Recent advances propose an offline mining mechanism to discover implicit novel instances, which exist in the base dataset, as auxiliary training samples to retrain a more powerful model. Nonetheless, the offline mined novel instances remain unchanged during retraining, thus hindering further improvements. A straightforward alternate adopts an online mining mechanism that employs an online teacher to mine implicit novel instances on the fly. However, the online teacher relies on a good initialization which is non-trivial in the scenarios of FSOD. To overcome the obstacles, we present Mining Implicit Novel Instances (MINI), a framework that unifies the offline mining mechanism and online mining mechanism with an adaptive mingling design. In offline mining, MINI leverages an offline discriminator to collaboratively mine implicit novel instances with a trained FSOD model. In online mining, MINI takes a teacher-student framework to simultaneously update the FSOD network and the mined implicit novel instances on the fly. In adaptive mingling, the offline and online mind implicit novel instances are adaptively combined, where the offline mined novel instances warm up the early training and the online mined novel instances gradually substitute the offline mined instances to further improve the performance. Extensive experiments on PASCAL VOC and MS-COCO datasets show MINI achieves new state-of-the-art performance on any shot and split of FSOD tasks. All code will be made available.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning

TL;DR: Mining Implicit Novel Instances for Few-Shot Object Detection