PXDepth preserves object boundaries with pixel-space depth modeling

A team led by Zhiyuan Yuan, with researchers including Guanying Chen and Ruimao Zhang, has released PXDepth, a discriminative monocular depth estimation model that tackles a persistent weakness in current deep learning approaches: losing fine-grained structures and object boundaries. The authors identify the root cause in the standard architecture combo of large-patch Vision Transformer (ViT) encoders and convolutional decoders, where coarse tokenization destroys pixel-level cues that upsampling cannot fully recover. To fix this, PXDepth separates global context modeling from pixel-level depth prediction. A large-patch ViT captures the overall scene, while a dedicated pixel-space predictor built from Context-Modulated Pixel Transformer blocks keeps high-resolution spatial representations throughout the entire depth estimation process.

This design lets PXDepth preserve fine structures and sharp edges without sacrificing global depth consistency. Across diverse zero-shot benchmarks, the model combines faithful local geometry with competitive global depth accuracy, all while remaining efficient at inference compared to other zero-shot depth estimators. The paper is available on arXiv (2608.16984) and the authors have released code and model weights, enabling immediate experimentation and integration into existing pipelines. For computer vision practitioners working on AR/VR, robotics, or 3D reconstruction, PXDepth offers a promising alternative to existing monocular depth models that often produce blurry, boundary-smudged depth maps. The work also highlights a broader architectural insight: decoupling global and local processing can lead to better performance on pixel-precise tasks.