Weakly Supervised Object Localization Using A Self-Training Approach

Abstract

近年物件定位與偵測在深度學習的研究上持續受到關注，許多物件定位的技術廣泛應用於我們的產業與生活中，為了達成可以在現實應用為目標，我們必須考量在現實應用中缺乏完整標註資料與標註成本的問題，且模型的準確率與泛化能力更是重要的指標。弱監督物件定位(Weakly supervised object localization)的研究，是在訓練集資料中缺乏位置級標籤，僅能使用圖像級標籤進行物件定位的工作。最近一個弱監督物件定位規範被提出，其包含一個具位置級標籤但少量的驗證集，供研究者調整超參以訓練出強健的定位模型。在本論文中，我們提出以半監督學習的技術應用於弱監督物件定位工作。我們獨立拆分定位模型與分類模型，可以解決少樣本的物件定位問題，並且避免分類影響定位的性能。我們提出的定位模型以半監督學習的方式進行訓練，首先使用非常少量有標籤的資料訓練定位模型，再利用其為未標註資料產生偽標籤。我們又提出一個偽標籤篩選演算法，從兩個不同的定位結果利用其互補特性，選擇高品質偽的標籤，並解決資料不平衡與樣本難易度差異的問題。最後我們將篩選後的偽標籤作為訓練資料再次訓練模型，配合使用訓練樣本預訓練的分類器進行分類與辨識。我們提出的方法可以有效利用弱標註的訓練資料，降低資料標註成本。我們所提出的方法不僅能有效提升模型的準確率，同時也將模型測試於不同的資料集中展現模型的泛化能力。

Object localization and detection have received increasing attention in computer vision and enable a wide range of applications. In the development of such deep learning techniques, we must consider the scale and the cost of data annotations, as well as the accuracy and the generalization ability of the developed model. Weakly supervised object localization deals with the lack of location-level labels in the training data---only image-level labels are available to train a model for inferring object locations. Recently, researchers found this problem ill-posed and propose a benchmark that involves a small amount of full supervisions for training a model. In this paper, we present a semi-supervised learning method to address the problem. In particular, we decompose this task into class-agnostic object detection and image classification, alleviating the need of samples to train a robust detector. The proposed localization model is developed via self-training: we use a small amount of data will full supervision to train the class-agnostic detector, and then use it to generate pseudo bounding boxes for data with weak supervision. Furthermore, we propose a selection algorithm to discover high-quality pseudo labels, along with a loss function to deal with data imbalance. We show that the proposed semi-supervised learning method is competitive to state-of-the-art methods. Moreover, it has the potential to generalize well on different datasets.