VOLO: Vision Outlooker for Visual Recognition.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Vision Transformers (ViTs) show promise in visual recognition but struggle with fine-level feature encoding, lagging behind CNNs when trained from scratch.
• Existing ViTs exhibit low training sample efficiency due to simple tokenization and limited feature richness from redundant backbone designs.

Purpose of the Study:

• To address the limitations of Vision Transformers in encoding fine-level features and improve their performance on visual recognition tasks.
• To introduce a novel, efficient, and generic architecture called Vision Outlooker (VOLO) that overcomes the drawbacks of current ViT designs.

Main Methods:

• Developed Vision Outlooker (VOLO) featuring a novel outlook attention mechanism for dynamic local feature aggregation in a sliding window manner.
• Outlook attention focuses on encoding fine-level features, unlike self-attention's global dependency modeling, offering improved memory efficiency and breaking computational bottlenecks.

Main Results:

• VOLO with 26.6M parameters achieved 84.2% top-1 accuracy on ImageNet-1K, surpassing T2T-ViT by 2.7% with comparable parameters.
• Scaling VOLO to 296M parameters resulted in 87.1% accuracy, setting a new record for ImageNet-1K classification.
• Pretrained VOLO models demonstrated superior performance on downstream tasks, including semantic segmentation.

Conclusions:

• Vision Outlooker (VOLO) offers a more efficient and effective approach to visual recognition by improving fine-level feature encoding.
• The outlook attention mechanism provides a scalable and memory-efficient alternative to self-attention, enabling state-of-the-art performance on benchmark datasets and downstream applications.