MViT improves classification in computer vision, highlighting better adaptability to complex structures.
Key Points
The aim is to improve Vision Transformers' ability to represent complex images by addressing limitations in spatial continuity and positional encoding.
Implemented a multi-order fractal mapping to optimize patch reordering.
Designed a dynamic partitioning template with a boundary compensation algorithm.
Integrated a period-aware positional encoding module with convolutional features.
Achieved a 0.52% increase in classification accuracy on CIFAR-100 compared to ViT-B/16.
Attained a 0.31% increase in accuracy on ImageNet-21k.
Showed improved metrics in PSNR and SSIM, demonstrating robust performance against rotation and variations.