Reliable evaluation of visual 3D reconstruction and depth estimation in vegetation-rich outdoor environments is limited by the scarcity of geometric ground-truth data and the cost of complete scanner surveys. We present a LiDARbased reference-generation pipeline for evaluating the zero-shot performance of feed-forward visual-geometry models alongside conventional structure-from-motion (SfM) methods. The pipeline reuses camera images, LiDAR scans, and reference poses from existing multi-sensor localization datasets to generate per-view depth maps and visible reference point clouds. To address visibility errors in accumulated LiDAR scans, we introduce the LargeDepth-Gap Visibility Score (LDGVS) for point-wise visibility filtering and automatic scale selection in Hidden Point Removal(HPR). Experiments on TartanAirV2 and TartanGround show that, compared with Poisson surface reconstruction, which is an approach suited to completed scan surveys, the combined HPR and LDGVS filtering retains approximately three times the point density at comparable absolute relative error. We also apply the pipeline to the BotanicGarden dataset to systematically compare depth estimation and 3D reconstruction methods. Feedforward methods produce dense depth predictions efficiently, but their 3D reconstructions can contain duplicated structures. Conventional SfM generally produces more coherent geometry, but its reconstructions remain sparse and more prone to failure in challenging vegetated scenes. The benchmark toolkit and datasets will be released.
Figures & tables
Fig. 1: Visualization of the reference data generated by the proposed pipeline in BotanicGarden. Top: the accumulated scene-level reference point cloud and the camera trajectory shown in red. The colored boxes indicate the two individual-view examples below. Bottom: for each example, the upper row shows the corresponding visibility-filtered point cloud in the camera frame, and the lower row shows the projected sparse reference depth image.
Dataset
GT Trajectory
Data Scale
SFU-Mountain [ 31 ]
Manually aligned locations
> 8 h, 150 GB
RELLIS-3D [ 32 ]
VoxNet poses
5 seq., 13.6k scans
Wild-Places [ 33 ]
Wildcat SLAM poses
8 seq., 63k submaps
TreeScope [ 34 ]
Faster-LIO poses
6 sites, > 1.8k labels
CitrusFarm [ 35 ]
GPS-RTK
7 seq., 1.7 h, 1.3 TB
ARD-VO [ 36 ]
GPS-RTK
11 sessions, > 1 TB
TABLE I: Multi-sensor localization datasets in vegetation-rich outdoor environments.
Fig. 2: Overview of the LiDAR-based reference-generation pipeline. In the “Deskew and denoise” step, the retained point cloud is shown in green and the raw point cloud in red.
Fig. 3: Examples of the visibility problem in single-scan and accumulated LiDAR point clouds. Background points are projected onto foreground tree regions, producing incorrect occlusion relationships.
Fig. 4: Examples of local depth histograms constructed from the 75 nearest projected LiDAR samples. Correct projections, such as the middle example, produce no distinct depth gap, whereas incorrect projections, such as the left and right examples, produce a clear gap.
Fig. 5: Effect of the HPR radius scale in close-range (top row) and far-range (bottom row) scenes. Results are shown for s=500 (left column) and s=5000 (right column). A scale that gives an acceptable visibility estimate at close range can retain many incorrect depth samples at long range, motivating automatic radius scale selection.
Method
AbsRel ↓
Rel τ↑
Abs ↓
Abs τ↑
Density ↑
Single scan (left view)
0.1138
0.7782
3.7065
0.4652
0.0232
Single scan (right view)
0.4024
0.7448
4.2277
0.4030
0.0216
Cumulation (baseline)
0.4016
0.6384
5.4760
0.3098
0.2524
Cumulation+CP
0.3214
0.6393
5.5628
0.3176
0.3171
PSR
0.1386
0.7336
3.2805
0.4297
0.0582
SGM
0.3093
0.5919
3.6819
0.2426
0.0065
TABLE II: Evaluation of visible depth on TartanAirV2 Data_easy and TartanGround Data_omni. The best value per metric is shown in bold and the second-best distinct value is underlined.
Fig. 6: Qualitative comparison of visible-depth projections for a TartanAirV2 forest scene. Panels (a) and (b) show the right-view RGB image and ground-truth depth, respectively; panels (c)–(l) show the projections generated by the selected comparison and proposed methods.
Method
Filter
Align
Seq.
Acc ↓
Acc τ↑
Comp ↓
Comp τ↑
CD ↓
F1 ↑
Density ↑
Runtime ↓
Feed-forward dense reconstruction
VGGT [ 17 ]
–
✓
10/10
0.3136
0.2719
0.6837
0.4106
0.9973
0.3236
1.0000
0.0508
VGGT [ 17 ]
P50
✓
10/10
0.1059
0.4587
1.4361
0.2845
1.5421
0.3486
0.5018
0.0503
VGGT- Ω [ 18 ]
–
✓
10/10
0.2753
0.3442
0.2171
0.4848
0.4924
0.4015
1.0000
0.0287
VGGT- Ω [ 18 ]
P50
✓
10/10
0.0639
0.6100
1.0890
0.3325
1.1529
0.4282
0.5000
0.0284
DA3 [ 19 ]
–
✓
10/10
0.3311
0.2462
0.3967
0.3998
0.7278
0.3036
1.0000
0.0539
TABLE III: Quantitative comparison of selected 3D reconstruction methods. Within each method block, the best value per column, except Density, is in bold and the second-best distinct value is underlined; The Align column indicates whether scene-level point-cloud alignment is applied before evaluation. In method labels, K denotes the calibrated-intrinsics-only configuration, and S denotes the calibrated-stereo configuration. All methods in the SfM block use calibrated intrinsics. The Filter column identifies confidence-P50 masking (P50); a dash indicates that this operation is not applied. The Seq. column reports the number of successfully evaluated sequences.
Fig. 7: Qualitative 3D reconstruction comparison on BotanicGarden_1006-03 . Columns show four fixed viewpoints of the same scene. The first row is the reference point cloud, followed by the feed-forward and SfM reconstruction methods. All methods use monocular input. VGGT, VGGT- Ω , DA3, MapAnything, and MASt3R-SfM are shown with confidence-P50 masking. All predictions are shown after alignment to the reference coordinate frame.
Method
Filter
Align
Abs ↓
Abs τ↑
AbsRel ↓
Rel τ↑
Density ↑
Runtime ↓
Monocular depth estimation
DA2 (Relative) [ 69 ]
–
✓
3.9086
0.0090
0.4332
0.0392
0.9410
0.0160
DA2 (Metric) [ 69 ]
–
×
11.3001
0.0001
2.0257
0.0041
1.0000
0.0157
DA2 (Metric) [ 69 ]
≤10 m
×
5.2689
0.0000
1.9307
0.0002
0.2693
0.0158
DA2 (Metric) [ 69 ]
–
✓
2.3071
0.0234
0.2497
0.0940
1.0000
0.0157
DA2 (Metric) [ 69 ]
≤10 m
✓
0.2159
0.1360
0.0761
0.3486
0.2693
0.0158
TABLE IV: Quantitative comparison of depth-estimation methods. Within each method block, the best value per column, except Density, is in bold and the second-best distinct value is underlined; The Align column indicates whether per-frame affine depth alignment is applied before evaluation. In method labels, K denotes camera intrinsics and P denotes camera poses; The Filter column identifies confidence-P50 masking (P50) or the range restriction ( ≤10 m); a dash indicates that neither operation is applied.
Fig. 8: Qualitative dense-depth comparison on BotanicGarden_1006-01 . Full Depth shows the complete valid prediction for assessing relative-depth discrimination, whereas Filtered Depth shows the subset retained by method-specific confidence or range filtering. 3D View 1 and 3D View 2 show two fixed views of the point clouds, back-projecting the retained depth maps, and accumulating them across the sequence using the camera poses associated with each configuration. The first row contains, from left to right, the undistorted input image, reference depth, and two views of the reference point cloud; the remaining rows show the monocular, stereo, and multi-view depth-estimation methods.
Method
AbsRel ↓
Rel τ↑
Abs ↓
Abs τ↑
Density ↑
Cumulation (baseline)
0.4016
0.6384
5.4760
0.3098
0.2524
LDGVSF threshold ablation
LDGVSF(0.5)
0.2268
0.6837
4.4518
0.3368
0.2148
LDGVSF(0.7)
0.1819
0.6918
4.3071
0.3477
0.1944
LDGVSF(0.9)
0.1478
0.6915
4.2159
0.3569
0.1684
AHPR target low-visibility ratio and LDGVSF threshold ablation
TABLE V: Ablation study of LDGVSF thresholds and AHPR target low-visibility ratios on TartanAirV2 Data_easy and TartanGround Data_omni. The percentage in AHPR denotes the target low-visibility ratio ρ⋆ . Within each ablation block, the best value per metric is shown in bold and the second-best distinct value is underlined.
Fig. 9: Qualitative comparison of the unfiltered cumulation baseline and AHPR results with and without LDGVSF-based filtering of low-visibility-score points.
Fig. 10: Qualitative comparison of visible-depth projections generated from Livox LiDAR data for BotanicGarden sequence 1018-13 . Panel (a) shows the right-view RGB image; panels (b)–(k) show the projections generated by the selected comparison and proposed methods.
Fig. 11: Reference image and per-pixel depth-error maps for the TartanAirV2 forest scene. Panel (a) shows the right-view RGB image, and panel (b) shows the aligned left-view single-scan error. Panels (c)–(l) show the right-view errors for the ten projections in Fig. 6 (c)–(l), in the same order. Errors are displayed with a jet color map (blue: low error; red: high error).
Fig. 12: Reference outputs for the ten BotanicGarden benchmark sequences. Columns 1–2 show two 3D views of the colorized reference point clouds; columns 3–4 show two colorized reference depth.
Feed-forward 3D priors enable high-quality reconstruction from image streams but often lack reliable metric scale and suffer from long-term trajectory drift. In contrast, LiDAR--inertial estimation provides metrically accurate motion but lacks the image-space consistency needed for high-fidelity reconstruction. We propose MIRAGE, which anchors feed-forward visual geometry to LiDAR--IMU metric motion and regulates online pose corrections according to correspondence support and cross-modal agreement, jointly enabling accurate metric pose estimation and high-fidelity Gaussian reconstruction. Experiments on Oxford Spires, M2DGR and INS show that MIRAGE achieves accurate metric trajectory estimation, outperforming COLMAP, a widely used but computationally expensive reference for vision-only benchmarks. Motivated by the scarcity of long-horizon monocular benchmarks for large-scale indoor reconstruction, we further build a multimodal data-collection platform and use MIRAGE to establish accurate 6-DoF reference trajectories for a new vision-only benchmark. We benchmark recent monocular large-scale reconstruction methods on this dataset, providing a systematic evaluation of their performance in long-horizon indoor scenes.
Xinzhe Wang, Changjian Jiang, Kaiwen Song +8
Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory · The University of Hong Kong +4
We introduce ReconSplat, a feed-forward model for 3D scene reconstruction that aims to address the longstanding trade-off between plausible view generation for unobserved regions and geometric consistency, providing both geometrically aligned novel views and sharp depth estimates. Our approach builds on 3D Gaussian splatting (3DGS) as an intermediate differentiable scene representation and integrates it with a multi-view latent diffusion model (MV-LDM) trained to act simultaneously as a refiner and an inpainter for appearance and scene geometry. We enforce geometric consistency by guiding the diffusion process with variational 3D latent features for appearance and geometry, encoded by the feed-forward 3DGS representation and rasterized to 2D latent space. ReconSplat produces both photorealistic novel views and accurate depth maps on real-world benchmarks, RealEstate10K and DL3DV-10K, outperforming existing methods in challenging extrapolation setups. Notably, ReconSplat allows the extrapolation of unseen and challenging viewpoints jointly with coherent and precise scene geometry.
Giuseppe Stracquadanio, Kevin Raj, Julia Grabinski +1
LiDAR point clouds acquired in underground environments exhibit severe geometric incompleteness due to occlusions and limited sensor viewpoints, making reliable point cloud completion challenging without large supervised datasets. We propose ZVeC, a zero-shot, instance-driven framework that reformulates scene-level completion as compositional object-level reconstruction. By decomposing a scene into semantic object instances, ZVeC reduces reconstruction ambiguity in cluttered environments while eliminating the need for scenario-specific training. Each segmented vehicle is completed independently using a depth- and 3D Gaussian-conditioned diffusion model that exploits generalized geometric priors before the reconstructed instances are recomposed into the original scene. To evaluate our approach, we construct a real-world dense LiDAR benchmark of underground parking environments. Experimental results demonstrate consistent improvements over representative scene-level baselines in both quantitative metrics and visual quality. The completed point cloud differs substantially from the measured input (average KL divergence ~ 2.1), yet reducing the input to only 1% of the original LiDAR measurements changes the completed reconstruction only marginally (KL divergence < 0.50). This demonstrates that ZVeC produces geometrically consistent completions even under extreme input sparsity.
Daisy Li, Kyle Gao, Quanyun Wu +3
University of Waterloo · Karlsruhe Institute of Technology