CMT: Convolutional Neural Network Meet Vision Transformers (CMT)

CMT implements a novel architecture that progressively transitions from convolutional feature extraction to transformer-based attention by using convolutional token embedding (CTE) blocks in early stages and multi-head self-attention in later stages. Early layers leverage 2D convolutions to capture local spatial patterns with inductive bias, while later layers apply transformer attention to learn global dependencies. This hybrid approach reduces computational complexity compared to pure ViT while maintaining spatial awareness through convolutional priors, using a staged fusion pattern where CNN features are tokenized before transformer processing.

CMT constructs multi-scale feature representations across different spatial resolutions using a pyramid structure where each stage outputs features at progressively coarser resolutions. Features from different scales are fused using attention mechanisms rather than simple concatenation, allowing the model to learn which scale-specific features are most relevant for the task. This attention-based fusion enables dynamic weighting of multi-scale information, improving performance on objects of varying sizes and improving robustness to scale variations in natural images.

CMT implements self-attention with spatial locality constraints by restricting attention computation to local windows rather than computing global attention over the entire feature map. This reduces attention complexity from O(N²) to O(N·W²) where W is the window size, enabling practical application of Transformers to high-resolution feature maps. The implementation uses shifted window attention patterns (similar to Swin Transformer) where windows are shifted between layers to enable cross-window information flow while maintaining computational efficiency.

CMT implements a hierarchical feature pyramid where spatial resolution decreases progressively through the network (224→112→56→28 pixels) while feature channel dimension increases correspondingly (64→128→256→512 channels). This design pattern, inherited from CNNs, maintains computational efficiency by reducing the spatial dimensions where expensive operations (like attention) are applied. The progressive reduction is achieved through strided convolutions or patch merging operations that combine adjacent spatial locations while expanding the feature representation capacity.

CMT provides a shared feature extraction backbone that can be adapted to different vision tasks (classification, detection, segmentation) through task-specific decoder heads. The backbone learns general-purpose visual representations through supervised or self-supervised pretraining, which are then fine-tuned or frozen for downstream tasks. This design enables efficient transfer learning and reduces the need to train separate models for different tasks, leveraging the hybrid CNN-Transformer architecture's ability to capture both local and global visual patterns useful across diverse applications.

CMT replaces Vision Transformer's linear patch embedding with learnable convolutional token embedding (CTE) blocks that use grouped convolutions to create tokens from image patches. Instead of flattening and projecting patches linearly, CTE applies multiple grouped convolution layers with progressively larger receptive fields to capture spatial structure within patches before tokenization. This approach preserves spatial relationships and local patterns within tokens, providing stronger inductive bias than linear projection while maintaining computational efficiency through grouped convolution implementations.