[Paper Review] Dual-attention Focused Module for Weakly Supervised Object Localization

Yukun Zhou, Zailiang Chen+4 more|arXiv (Cornell University)|Sep 11, 2019

Advanced Neural Network Applications30 references4 citations

TL;DR

This paper proposes the Dual-attention Focused Module (DFM), a novel attention mechanism for weakly supervised object localization that jointly models spatial and channel-wise attention to enhance feature representation. By integrating a position branch and a channel branch with focused matrix regularization and feature fusion, DFM improves localization accuracy, achieving state-of-the-art performance on ILSVRC 2016 and CUB-200-2011 benchmarks.

ABSTRACT

The research on recognizing the most discriminative regions provides referential information for weakly supervised object localization with only image-level annotations. However, the most discriminative regions usually conceal the other parts of the object, thereby impeding entire object recognition and localization. To tackle this problem, the Dual-attention Focused Module (DFM) is proposed to enhance object localization performance. Specifically, we present a dual attention module for information fusion, consisting of a position branch and a channel one. In each branch, the input feature map is deduced into an enhancement map and a mask map, thereby highlighting the most discriminative parts or hiding them. For the position mask map, we introduce a focused matrix to enhance it, which utilizes the principle that the pixels of an object are continuous. Between these two branches, the enhancement map is integrated with the mask map, aiming at partially compensating the lost information and diversifies the features. With the dual-attention module and focused matrix, the entire object region could be precisely recognized with implicit information. We demonstrate outperforming results of DFM in experiments. In particular, DFM achieves state-of-the-art performance in localization accuracy in ILSVRC 2016 and CUB-200-2011.

Motivation & Objective

• To address the challenge of incomplete object localization in weakly supervised object detection, where only image-level labels are available.
• To improve feature representation by focusing on discriminative regions while preserving contextual information from the full object.
• To mitigate the problem of hidden object parts being obscured by highly discriminative patches during localization.
• To enhance localization performance through a novel attention mechanism that fuses spatial and channel-wise features effectively.

Proposed method

• The Dual-attention Focused Module (DFM) consists of two parallel branches: a position branch and a channel branch for spatial and channel-wise attention, respectively.
• Each branch generates an enhancement map and a mask map from the input feature map to either highlight or suppress specific features.
• A focused matrix is introduced in the position branch to enforce spatial continuity of object pixels, improving localization consistency.
• The enhancement map from each branch is fused with the corresponding mask map to recover lost features and diversify representations.
• The fused features are used to refine the final localization map, improving detection accuracy with minimal computational overhead.
• The module is integrated into a standard weakly supervised object detection framework, trained end-to-end with cross-entropy loss.

Experimental results

Research questions

• RQ1 How can attention mechanisms be designed to improve localization by focusing on discriminative parts without losing full object context?
• RQ2 Can spatial continuity constraints enhance the robustness of object localization in weakly supervised settings?
• RQ3 To what extent can joint modeling of spatial and channel-wise attention improve feature representation for weakly supervised object localization?
• RQ4 Does the proposed dual-attention module with focused matrix regularization outperform existing attention mechanisms on standard benchmarks?

Key findings

• The DFM achieves state-of-the-art mean Intersection over Union (mIoU) performance on the ILSVRC 2016 dataset, outperforming prior methods.
• On the CUB-200-2011 benchmark, DFM achieves the highest localization accuracy among all compared weakly supervised methods.
• Ablation studies confirm that both the dual-attention mechanism and the focused matrix contribute significantly to performance gains.
• The method demonstrates robustness across diverse object categories and complex backgrounds, indicating strong generalization capability.