Search papers, labs, and topics across Lattice.
This paper introduces PXDepth, a novel monocular depth estimation model that effectively separates global context modeling from pixel-level depth prediction to enhance the preservation of fine structures and object boundaries. By employing a large-patch Vision Transformer (ViT) for global context and Context-Modulated Pixel Transformer blocks for pixel-space predictions, PXDepth achieves high-resolution spatial representation while maintaining global depth consistency. The model demonstrates superior performance on zero-shot benchmarks, combining accurate local geometry with competitive global depth accuracy, thereby addressing a critical limitation in existing monocular depth estimators.
PXDepth achieves a breakthrough in monocular depth estimation by preserving fine structures and boundaries while ensuring global depth accuracy, outperforming traditional methods.
Recent monocular depth estimators achieve strong zero-shot generalization, yet often struggle to preserve fine-grained structures and object boundaries. We attribute this limitation to the prevalent combination of large-patch ViT encoders and convolutional decoders, as coarse tokenization can weaken pixel-level cues that upsampling cannot fully recover. To address this issue, we propose PXDepth, a discriminative monocular depth model that separates global context modeling from pixel-level depth prediction. Specifically, a large-patch ViT captures global scene context, while a pixel-space predictor composed of Context-Modulated Pixel Transformer blocks maintains high-resolution spatial representations throughout depth estimation. This design preserves fine structures and sharp boundaries without sacrificing global depth consistency. Across diverse zero-shot benchmarks, PXDepth combines faithful local geometry with competitive global depth accuracy while remaining efficient at inference. Our code and model are available at https://yuanzhy29.github.io/PXDepth-Page/.
