Attention-Based Dropout Layer for Weakly Supervised Object Localization

Abstract

Weakly Supervised Object Localization (WSOL) techniques learn the object location only using image-level labels, without location annotations. A common limitation for these techniques is that they cover only the most discriminative part of the object, not the entire object. To address this problem, we propose an Attention-based Dropout Layer (ADL), which utilizes the self-attention mechanism to process the feature maps of the model. The proposed method is composed of two key components: 1) hiding the most discriminative part from the model for capturing the integral extent of object, and 2) highlighting the informative region for improving the recognition power of the model. Based on extensive experiments, we demonstrate that the proposed method is effective to improve the accuracy of WSOL, achieving a new state-of-the-art localization accuracy in CUB-200-2011 dataset. We also show that the proposed method is much more efficient in terms of both parameter and computation overheads than existing techniques.

Dual-Gradients Localization Framework for Weakly Supervised Object Localization

“…Previous works probe various technologies to discover integral object regions, consist of erasing [2,34], data argumentation [20,31], fusion maps [29], divergent activation [28], self-produced guidance learning [35]. Singh et al [20] proposed a data argumentation approach Hide and Seek (Has) to seek more object regions by randomly hiding an image with patches.…”

“…ACoL utilized the first classifier to localize the discriminative regions, and compelled second classifier to discover complementary object regions. Similarly, Junsuk et al [2] proposed Attention-based Dropout Layer (ADL) that erased high-scoring regions and forced network to have attention to low-scoring regions within feature maps. SPG [35] leveraged high confident object regions as auxiliary supervision to force classification model to learn more object regions.…”

“…We evaluate the proposed DGL framework on following popular CNNs: VGGnet [19], InceptionV3 [21], Resnet50 [4], Resnet50 [4] with SE block [5], and modify those backbones as following baselines [2,35,36]. For VGG16 network, we remove the layers after conv5_3, and add a convolutional layer named 'conv5_4', followed by a GAP layer and a FC layer with softmax function.…”

“…Weakly supervised learning has made remarkable progress recently. Existing methods have been successfully applied on object localization [2,9,17,20,28,31,[33][34][35][36][37], object detection [3,10,23,30] and segmentation [1,6,8,13,[24][25][26]32]. Weakly Supervised Object Localization (WSOL) refers to mining information of object locations by only using image-level annotations.…”

“…then performed CAM to localize the entire object regions. Some previous works [2,20,24,31,34] proposed erasing-based approaches to discover more regions. Other works [28,29,35] exploited spatial correlation information in feature maps to highlight the entire object regions.…”