Multimodal Diffusion-Based Depth Estimation Framework with Wi-Fi and Vision

Engineering notes

By leveraging the iterative refinement capabilities of diffusion models, WiViD achieves high-precision depth predictions. Real-world experiments demonstrate that WiViD significantly outperforms state-of-the-art monocular depth estimation methods, reducing Absolute Relative Error (ARE) and Square Relative Error (SRE) by 33.3% and 9.5%, respectively, highlighting its superior accuracy and robustness.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Depth estimation is essential for applications such as autonomous driving, robotic navigation, and augmented reality. While LiDAR-based and mmWave-based approaches offer high accuracy, they come with substantial deployment costs, whereas monocular vision methods often struggle with precision. To overcome these challenges, we propose WiViD, a novel diffusion-based depth estimation framework that integrates commercial Wi-Fi and visual data. By leveraging the iterative refinement capabilities of diffusion models, WiViD achieves high-precision depth predictions. WiViD features two key encoding modules: the Complex-Valued CSI Encoder (CCE), which extracts rich spatio-temporal features from Wi-Fi signals, and the Residual Image Encoder (RIE), which processes visual data. Additionally, we introduce the Multi-Modal Alignment Bidirectional Cross-Attention (MABA) mechanism to enhance cross-modal feature integration. Real-world experiments demonstrate that WiViD significantly outperforms state-of-the-art monocular depth estimation methods, reducing Absolute Relative Error (ARE) and Square Relative Error (SRE) by 33.3% and 9.5%, respectively, highlighting its superior accuracy and robustness.

Links and sources

Comments