Interweaving Insights: High-Order Feature Interaction for Fine-Grained Visual Recognition

• DOI: 10.1007/s11263-024-02260-y
• Publication Date: 2024-10-20T07:01:54Z
• Authors: Arindam Sikdar; Yonghuai Liu; Siddhardha Kedari...; Yitian Zhao; Amr Ahmed; Ardhendu Behera
• Keywords: Feature (linguistics); Feature Learning; Representation; Benchmark (surveying); Code (set theory); Abstraction

Abstract This paper presents a novel approach for Fine-Grained Visual Classification (FGVC) by exploring Graph Neural Networks (GNNs) to facilitate high-order feature interactions, with specific focus on constructing both inter- and intra-region graphs. Unlike previous FGVC techniques that often isolate global local features, our method combines features seamlessly during learning via Inter-region graphs capture long-range dependencies recognize patterns, while delve into finer details within regions of an object high-dimensional convolutional features. A key innovation is the use shared GNNs attention mechanism coupled Approximate Personalized Propagation Predictions (APPNP) message-passing algorithm, enhancing information propagation efficiency better discriminability simplifying model architecture computational efficiency. Additionally, introduction residual connections improves performance training stability. Comprehensive experiments showcase state-of-the-art results benchmark datasets, affirming efficacy approach. work underscores potential GNN in modeling high-level distinguishing it from methods typically singular aspects representation. Our source code available at https://github.com/Arindam-1991/I2-HOFI .