DPCT: Efficient High-Resolution Depth Prediction via Cross-Covariance Attention Transformers

Depth prediction tasks hold significant applications in fields like autonomous driving and augmented reality. However, existing Vision Transformer (ViT)-based models face efficiency bottlenecks in high-resolution scenarios due to the quadratic computational complexity of their self-attention mechanisms. This paper proposes an efficient depth prediction framework, DPCT (Depth Prediction Cross-Covariance Transformer), which replaces the ViT encoder with a Cross-Covariance Image Transformer (XCiT) to leverage its linear-complexity Cross Covariance Attention (XCA) mechanism, significantly reducing computational overhead. Comprehensive evaluations on indoor and outdoor datasets demonstrate the framework's strong generalization capability across diverse scenarios. Experimental results indicate that DPCT achieves competitive performance while substantially reducing model complexity and inference latency compared to conventional Transformer-based architectures, approaching real-time processing speeds. Further analysis reveals the model maintains robust efficiency across multiple input resolutions, exhibiting practical scalability for high resolution applications. This study provides a novel solution for deploying depth prediction models in resource-constrained scenarios, validates the effectiveness of channel-wise attention mechanisms in vision tasks, and demonstrates deployment potential through cross-dataset evaluations and multi-resolution performance verification.