通過對比式學習改進半監督物件偵測

Abstract

本篇論文中，我們研究了如何使用對比式學習幫助半監督物件偵測模型更好的學習。在半監督物件偵測中，使用無標注資料的流程相當於自監督學習的概念，因此我們參考自監督學習常用的對比式學習架構，提出了一種表徵等級的對比式學習。在半監督物件偵測中，常使用偽標籤法來進行學習，其核心概念是強、弱增強的圖像要有一致的預測，而對比式學習也有類似的假設，只是對於圖像的增強並沒有強、弱的要求。所以我們基於使用偽標籤法的半監督物件偵測模型，加上對比式學習的架構幫助模型學習。該方法通過計算無標注圖像的兩種視覺增強生成特徵的相似度，來改善半監督物件偵測的效果。實驗結果顯示，我們的方法在使用MS-COCO的1%、5%、10%有標注資料時，mAP結果分別提升了2.92%、1.88%、0.99%。這證明了對比式學習在半監督物件偵測中可以在特徵層面學習到額外的資訊，並且基於原本的半監督物件流程，加上對比式學習的流程並不需要增加太多額外的計算。我們期望這項研究能為半監督學習及物件偵測的未來研究提供新的思路和方向。In this paper, we investigate how to use contrastive learning to enhance semi-supervised object detection. In semi-supervised object detection, the process of utilizing unlabeled data is similar to the concept of self-supervised learning. Hence, we refer to the commonly used contrastive learning framework in self-supervised learning and propose a representation-level contrastive learning method.In semi-supervised object detection, pseudo-labeling is commonly used for self-learning. The core concept is that strongly and weakly augmented images should have consistent predictions. Contrastive learning has a similar assumption, except that there is no strong or weak requirement for image augmentation. Therefore, based on the semi-supervised object detection model with pseudo-labeling, we apply a contrastive learning framework to aid model learning with more feature-level information. This method improves the performance of semi-supervised object detection by computing the similarity of features generated by two different augmentations of unlabeled images. The experimental results show that our method improves the mAPs results by 2.92%, 1.88%, and 0.99% when using 1%, 5%, and 10% of the labeled data from MS-COCO, respectively.This demonstrates that contrastive learning can learn additional information at the feature level in semi-supervised object detection, and adding a contrastive learning process to the original semi-supervised object process does not require much additional computation. We hope that this research can provide new insights and directions for future research in semi-supervised learning and object detection.