Researchers have proposed a new approach to monocular depth estimation called PXDepth. The model separates global context modeling from pixel-level depth prediction using a large-patch Vision Transformer (ViT) encoder and a Context-Modulated Pixel Transformer block predictor. This design allows for preservation of fine structures and sharp boundaries without sacrificing global depth consistency. According to the authors, PXDepth achieves competitive results on zero-shot benchmarks while remaining efficient at inference. --- Why it matters: This matters because existing monocular depth estimators often struggle with preserving fine-grained structures and object boundaries, which is crucial for many applications such as robotics and autonomous vehicles. Source: https://arxiv.org/abs/2608.16984