Sun Bo ; Li Banghuai ; Cai Shengcai ; Yuan Ye ; Zhang Chi

Abstract

Emerging interests have been brought to recognize previously unseen objectsgiven very few training examples, known as few-shot object detection (FSOD).Recent researches demonstrate that good feature embedding is the key to reachfavorable few-shot learning performance. We observe object proposals withdifferent Intersection-of-Union (IoU) scores are analogous to the intra-imageaugmentation used in contrastive approaches. And we exploit this analogy andincorporate supervised contrastive learning to achieve more robust objectsrepresentations in FSOD. We present Few-Shot object detection via Contrastiveproposals Encoding (FSCE), a simple yet effective approach to learningcontrastive-aware object proposal encodings that facilitate the classificationof detected objects. We notice the degradation of average precision (AP) forrare objects mainly comes from misclassifying novel instances as confusableclasses. And we ease the misclassification issues by promoting instance levelintra-class compactness and inter-class variance via our contrastive proposalencoding loss (CPE loss). Our design outperforms current state-of-the-art worksin any shot and all data splits, with up to +8.8% on standard benchmark PASCALVOC and +2.7% on challenging COCO benchmark. Code is available at: https://github.com/MegviiDetection/FSCE