A2FA: Adversarial Angular Feature Alignment for Multimodal Indoor Object Recognition

Abstract

In recent years, object recognition has attracted significant attention as a fundamental task for scene understanding in commercial robotics, particularly within consumer electronic products such as smart home assistants and service robots. In indoor environments, RGB images captured by conventional cameras are prone to variations in lighting conditions and occlusions. In contrast, depth images capture 3-D structural information that helps reduce occlusion effects and handle low-light conditions, leading to more consistent recognition performance. While recent advancements in RGB-Depth (RGB-D) fusion have improved environmental adaptability, two key challenges persist: preserving fine-grained local features in complex environments and achieving robust cross-domain generalization from synthetic data to real-world scenarios. To address the aforementioned challenges, this paper proposes a cross-domain RGB-D indoor object recognition method based on a convolutional neural network (CNN) architecture. By incorporating depth image features, the proposed approach effectively reduces the influence of visual noise present in RGB images. A novel proxy task—angle prediction via circular cropping—is introduced to guide the model in attending to local feature discrepancies between the two modalities. In addition, a domain classifier proxy task, formulated using adversarial learning, is designed to improve the generalization ability of the feature extractor and mitigate the synthetic-to-real domain shift. Experimental results on multiple public datasets demonstrate that the proposed method achieves superior recognition accuracy compared to existing approaches.