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

Zhiyuan Yuan; Guanying Chen; Lingteng Qiu; Ruimao Zhang; Shuguang Cui; Xiaochun Cao

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

Zhiyuan Yuan¹ Guanying Chen^1* Lingteng Qiu³ Ruimao Zhang¹ Shuguang Cui^3,2 Xiaochun Cao¹

¹Sun Yat-sen University ²Shenzhen-FNii ³CUHKSZ

^*Corresponding author

TL;DR: PXDepth separates global context encoding from pixel-space depth prediction, using a Global Context Encoder and a Pixel-Space Depth Predictor.

Qualitative Visualization

PXDepth Ours

Loads when this section is visible

Baseline

Loads when this section is visible

Rotate left drag Zoom wheel or pinch Pan right drag or Shift-drag Focus double click Adjust align one view, then Done Reset R or button

Choose a scene Street II 11 / 14

Hover for a larger RGB preview

Overview

PXDepth teaser comparing indoor, outdoor, and wild-scene point clouds alongside global accuracy, boundary CD, and latency results.

Comparison To Existing Architecture

Comparison of convolutional decoding, multi-step diffusion decoding, and PXDepth pixel-space prediction.

Discriminative models. A large-patch ViT captures global context, while a convolutional decoder upsamples its low-resolution features in a single feed-forward pass. This is efficient, but coarse tokenization weakens high-frequency cues that upsampling may not fully recover, leading to distorted local structures.

PPD. Pixel-space diffusion, semantic guidance, and a cascaded DiT preserve high-frequency information and improve fine structures. However, multi-step denoising repeatedly evaluates the network, resulting in high latency.

PXDepth. A Global Context Encoder supplies scene-level geometry, while a CM-PiT-based Pixel-Space Depth Predictor maintains pixel-space features with coarse-to-fine compaction. This single-pass design preserves local structure while maintaining low latency.

Qualitative Comparison

Qualitative point-cloud comparison between MoGe-2, MDA, PPD, InfiniDepth, and PXDepth across indoor and outdoor scenes.

Quantitative Comparison

Category	Method	MoGe Benchmark		MDA Benchmark			Latency ↓
Category	Method	Rel ↓	δ₁ ↑	Rel ↓	δ₁ ↑	Boundary CD ↓	Latency ↓
Global Prediction	DA V2	8.0	93.0	4.7	96.5	112.8	55.3
	DepthPro	6.7	93.6	3.9	97.0	106.4	206.8
	MoGe-2	5.1	95.6	3.5	96.9	100.3	32.9
Structure Aware	InfiniDepth	7.2	94.1	4.7	96.7	98.9	65.7
	PPD	7.5	92.6	4.4	96.5	93.7	209.8
	MDA	8.2	90.5	3.6	96.7	72.8	267.1
	PXDepth	5.8	94.8	3.6	97.0	70.6	56.4

Rel and δ₁ measure global depth accuracy, while Boundary CD measures boundary sharpness. Boundary CD is reported in millimeters (mm). Latency is measured per image at 518 × 518 on an RTX 5880 GPU and reported in milliseconds (ms).

BibTeX

Coming soon