ReViT: Enhancing Vision Transformers Feature Diversity with Attention Residual Connections

TL;DR

The paper tackles the problem of feature collapse in Vision Transformer self-attention, which tends to erase useful low-level features as depth increases. It introduces residual attention learning via attention residual connections to preserve and diversify features across layers, thereby enhancing robustness and recognition accuracy. Empirical results demonstrate improvements on five image classification benchmarks—ImageNet1k, CIFAR10, CIFAR100, Oxford Flowers-102, and Oxford-IIIT Pet—while COCO2017 experiments show that incorporating the approach into spatial-aware transformer models yields richer semantic and spatial relationships for object detection and instance segmentation. The proposed method offers a practical route to strengthen ViT representations and enable better cross-task performance in detection and segmentation scenarios.

Abstract

Vision Transformer (ViT) self-attention mechanism is characterized by feature collapse in deeper layers, resulting in the vanishing of low-level visual features. However, such features can be helpful to accurately represent and identify elements within an image and increase the accuracy and robustness of vision-based recognition systems. Following this rationale, we propose a novel residual attention learning method for improving ViT-based architectures, increasing their visual feature diversity and model robustness. In this way, the proposed network can capture and preserve significant low-level features, providing more details about the elements within the scene being analyzed. The effectiveness and robustness of the presented method are evaluated on five image classification benchmarks, including ImageNet1k, CIFAR10, CIFAR100, Oxford Flowers-102, and Oxford-IIIT Pet, achieving improved performances. Additionally, experiments on the COCO2017 dataset show that the devised approach discovers and incorporates semantic and spatial relationships for object detection and instance segmentation when implemented into spatial-aware transformer models.