MoGe-3

Results

💡Tips

Zoom - Scroll mouse wheel or pinch

Rotate - Left drag

Pan - Right drag

Pick a point - Left click

Pick two points and measure the distance

Methodology

Self-Guided Sparse 3D Refinement (SSR). We lift the coarse point map from a base model onto a sparse voxel shell and refine the depth via lightweight sparse 3D convolutions. Operating in 3D space rather than the image plane avoids feature mixing across discontinuities between depth layers, recovering fine-detailed geometry that 2D decoders typically over-smooth.

MoGe-3 base model and Self-Guided Sparse 3D Refiner architecture

Comparison to Other Methods

MoGe-3 (Ours)

💡Tips

Zoom - Scroll mouse wheel or pinch

Rotate - Left drag

Pan - Right drag

Switch baselines from the dropdown above

Quantitative Comparison

Zero-shot averages across multiple benchmarks. Lower Rel is better; higher δ and F1 are better.

State-of-the-art fidelity. Across 9 zero-shot benchmarks, SSR sets the best global and local accuracy, with the largest gains on strict fine-detail metrics (δ_0.01). It matches Depth Pro's boundary sharpness at much lower resolution, while yielding far cleaner 3D structure.

Method	Local				Boundary	Global
	Depth		Point		F1↑	Relative Depth		Relative Point		Metric Depth
	Rel^d↓	δ^d_0.01↑	Rel^p↓	δ^p_0.01↑	F1↑	Rel^d↓	δ^d₁↑	Rel^p↓	δ^p₁↑	Rel^d↓	δ^d₁↑
Depth Pro	4.15	46.9	3.01	45.2	16.3	8.10	91.3	11.1	87.6	29.0	54.0
UniDepth V2	5.60	45.3	4.60	42.6	14.5	6.98	92.8	10.2	88.6	23.7	71.2
Depth Anything 3	10.5	41.8	7.39	44.7	6.61	7.89	91.4	10.6	87.5	17.4	70.1
UniK3D	5.52	46.2	4.37	45.0	15.1	6.95	92.8	9.63	89.6	17.9	77.8
Pixel-Perfect Depth	5.24	39.8	4.39	39.1	12.7	7.17	92.5	20.3	69.8	–	–
InfiniDepth	5.82	43.2	4.29	43.7	19.3	8.15	91.1	11.6	87.6	–	–
MoGe-2	3.81	47.1	3.19	46.6	15.6	5.98	94.1	8.73	90.4	15.6	77.3
Ours (ViT-L without refinement)	3.60	48.89	2.93	50.26	15.62	5.56	94.67	7.94	91.97	14.74	80.97
Ours (ViT-L)	3.76	52.5	2.79	55.9	16.0	5.52	94.7	7.91	92.0	15.0	82.7
Ours (ViT-G)	3.66	54.1	2.80	56.8	17.3	4.80	95.6	7.43	92.4	15.8	82.7

Takeaways

2D image-space decoders mix features across depth discontinuities, blurring boundaries and thin structures. SSR instead refines geometry directly in 3D, preserving surface separation and recovering finer detail.

3D beats 2D refinement

Our Self-Guided Sparse 3D Refiner (SSR) discretizes the coarse point map into a thin sparse voxel shell and refines it with sparse 3D convolutions. Because occluding and occluded surfaces fall into separate voxels, cross-boundary feature bleeding is avoided — something a parameter-matched 2D U-Net cannot achieve.

Depth-only iterative updates

A factorized representation (u, v, log Z) keeps the image-axis coordinates fixed and lets the refiner predict only log-depth residuals. Each iteration re-voxelizes the updated geometry, progressively sharpening surface boundaries in a self-guided loop.

Efficient & robust

With K=0 the model matches the MoGe-2 base runtime; the 3-step refiner adds only moderate overhead (121 ms). Trained with K=3, it generalizes to K=7 at test time without degradation, and scales further with a ViT-G backbone.

We believe the SSR module can serve as a general-purpose dense geometry prediction head and a modular add-on for broader geometry foundation models.

BibTeX

@misc{kong2026finedetailmonoculargeometryestimation,
      title={Fine-Detail Monocular Geometry Estimation with Self-Guided Sparse Volumetric Refinement}, 
      author={Lingyu Kong and Ruicheng Li and Ruicheng Wang and Sicheng Xu and Chengtang Yao and Jianfeng Xiang and Jiaolong Yang},
      year={2026},
      eprint={2607.17967},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2607.17967}, 
}

MoGe-3: Fine-Detail Monocular Geometry Estimation with Self-Guided Sparse Volumetric Refinement

MoGe-3 pursues high-fidelity geometry details, using Self-Guided Sparse 3D Refinement (SSR) to recover fine structures from a single image.