Recombining Vision Transformer Architecture for Fine-Grained Visual Categorization

Abstract

Fine-grained visual categorization (FGVC) is a challenging task in the image analysis field which requires comprehensive discriminative feature extraction and representation. To get around this problem, previous works focus on designing complex modules, the so-called necks and heads, over simple backbones, while bringing a huge computational burden. In this paper, we bring a new insight: Vision Transformer itself is an all-in-one FGVC framework that consists of basic Backbone for feature extraction, Neck for further feature enhancement and Head for selecting discriminative feature. We delve into the feature extraction and representation pattern of ViT for FGVC and empirically show that simply recombining the original ViT structure to leverage multi-level semantic representation without introducing any other parameters is able to achieve higher performance. Under such insight, we proposed RecViT, a simple recombination and modification of original ViT, which can capture multi-level semantic features and facilitate fine-grained recognition. In RecViT, the deep layers of the original ViT are served as Head, a few middle layers as Neck and shallow layers as Backbone. In addition, we adopt an optional Feature Processing Module to enhance discriminative feature representation at each semantic level and align them for final recognition. With the above simple modifications, RecViT obtains significant improvement in accuracy in FGVC benchmarks: CUB-200-2011, Stanford Cars and Stanford Dogs.