PXDepth: Pixel-Space Modeling for Structure Preserving Monocular Depth Estimation

Details

Yuan et al. identify a specific failure mode in modern monocular depth estimators: the dominant large-patch ViT encoder combined with convolutional decoder architecture creates an information bottleneck. Coarse tokenization loses fine-grained pixel cues at the encoder stage, and subsequent upsampling in the decoder cannot fully recover sharp object boundaries or fine structural detail. PXDepth decouples these concerns by running two parallel streams: a large-patch ViT handles global scene context efficiently, while a pixel-space predictor using Context-Modulated Pixel Transformer (CMPT) blocks maintains full-resolution spatial representations throughout the depth estimation process. The CMPT blocks allow the pixel-space stream to incorporate global context from the ViT while preserving local resolution. Zero-shot benchmark results across diverse datasets show PXDepth combines faithful local geometry (fine structures and sharp boundaries) with competitive global depth accuracy while remaining computationally efficient at inference. Code and model weights are publicly released, making the method immediately testable. Applications where fine boundary accuracy matters — robotic manipulation, AR/VR depth integration, autonomous navigation in cluttered environments — stand to benefit most from the improved structure preservation.