Research map / 3D vision / Depth and stereo

SurGe: Improved Surface Geometry in Point Maps

NeurIPS 2026 · 3D vision › Depth and stereo · approach: Foundation models for monocular geometry · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SurGe, a monocular point-map model that improves local surface geometry via a new point map normal metric, a depth-normalized point gradient matching loss, and a Neighborhood Attention Decoder.

Problem

Feedforward point map models estimate global 3D geometry well but produce inaccurate local surface geometry (bent thin structures, surface noise) that standard pointwise metrics barely reflect.

Approach

Evaluate local surface orientation via point-map normals (MAE), supervise depth-normalized 3D finite differences with a pairwise scale-invariant gradient matching loss adapted to point maps, and replace convolutional decoders with a multi-stage Neighborhood Attention decoder.

Key results

• SurGe achieves the lowest local AbsRel on all five evaluated datasets (e.g., ETH3D 2.66% vs 3.16% for MoGe) and best average rank 1.62 on global point map AbsRel vs 2.25 for MoGe.
• Point map normal MAE improves over MoGe on all five datasets (e.g., ETH3D 18.3° vs 19.5°, Sintel 24.5° vs 25.2°).
• In the decoder ablation, NAD beats ConvStack-L on all local evaluations and 7 of 8 global datasets, but the decoder alone is 2.28x slower than ConvStack-L.

Surprising findings

• Log-depth gradient matching improves local geometry but harms global point map accuracy, while the proposed point gradient loss improves both.
• Removing LayerNorm from the NAD blocks improves accuracy but occasionally causes irrecoverable training instability.

Limitations

• NAD increases inference cost (decoder 2.28x slower than ConvStack-L at 512x512).
• Local artifacts remain in complex scenes with many high-frequency geometry changes; floaters around occlusion boundaries persist.
• Real-world SfM labels can suppress fine structures and let moving foreground objects bleed into background geometry.

Experimental setup

Datasets: Argoverse2 ARKitScenes BlendedMVS G-Objaverse GTA-SfM Hypersim IRS Ken Burns MatrixCity MegaDepth MidAir MVS-Synth

Benchmarks: NYUv2 KITTI ETH3D iBims-1 GSO Sintel DDAD DIODE HAMMER Spring

Metrics: AbsRel global point map AbsRel local point map point map normal mean angular error affine-invariant depth AbsRel delta_1 delta_0.05 depth boundary F1 average rank

Baselines: MoGe-2 MoGe InfiniDepth PPD VGGT VGGT-Ω Depth Anything 3 π^3 UniDepthV2 Depth Pro MapAnything DUSt3R

Builds on: MoGe DUSt3R VGGT log-depth gradient matching loss Neighborhood Attention DINOv2

Closest papers by meaning

• MoGe-2: Accurate Monocular Geometry with Metric Scale and Sharp Details (NeurIPS 2025 · similarity 0.90)
• Fine-Detail Monocular Geometry Estimation with Self-Guided Sparse Volumetric Refinement (NeurIPS 2026 · similarity 0.88)
• MoGe: Unlocking Accurate Monocular Geometry Estimation for Open-Domain Images with Optimal Training Supervision (CVPR 2025 · similarity 0.88)
• Depth Pro: Sharp Monocular Metric Depth in Less Than a Second (ICLR 2025 · similarity 0.86)
• PLATYPUS: Progressive Local Surface Estimator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2025 · similarity 0.85)
• Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · similarity 0.85)