Lens3D: What Visual Evidence Do 3D Feed-Forward Foundation Models Rely On?

1. Research motivation

Classical 3D reconstruction methods, such as Structure-from-Motion (SfM), rely on feature matching and two-view correspondence to recover camera poses and 3D structure. Their performance is directly tied to the availability and reliability of this geometric evidence.

My previous work on visual localization and reconstruction relied heavily on local features, geometric correspondences, and robust pose estimation. This motivated me to investigate whether modern feed-forward 3D foundation models depend on similar geometric evidence, despite learning reconstruction end-to-end without explicit featurematching.

In contrast, recent feed-forward 3D foundation models, such as VGGT, DA3, and AMB3R, predict geometry directly from images, without requiring an explicit matching and optimization pipeline. Their impressive reconstruction capabilities raise a fundamental question:

How sensitive are feed-forward 3D reconstruction models to the loss of geometric evidence? Which types of evidence matter most?

2. Evaluation Methodology

We investigate the sensitivity of feed-forward 3D reconstruction models, by hiding visual evidence in the input images and measuring the impact on reconstruction performance.

• Edges and Corners: Targets distinctive local structures using Harris corner and Canny edge responses to rank image patches.

• Geometric Correspondences: Targets cross-view matches detected with ALIKED and LightGlue and ranked by match density, and verified by fundamental-matrix RANSAC.

• Random (Low-evidence): Randomly masks regions with weak edge, corner, and correspondence evidence, mostly low-texture areas.

These three strategies probe different sources of geometric evidence while keeping the masking budget fixed. The low-evidence condition provides an area-matched reference, although differences in spatial and texture distributions may also influence reconstruction quality.

We evaluate four notable models: VGGT-1B, AMB3R, DA3, and VGGT-Omega. Performance is measured through 3D reconstruction accuracy, completeness, and camera pose accuracy.

3. Dataset Generation

We use ETH3D, comprising 104 multi-view samples, with images processed at a resolution of 518 × 392. Masking is applied using 14 × 14 pixel patches. For each masking strategy, we replace 10% or 25% of the image area with mid-gray pixels.

We generate 6 different masked datasets, and one clean baseline dataset without any masking.

Download ETH3D

Download the original ETH3D images and ground truth with scripts/download_rmvd.sh (~55 GB from Livioni/eth3d_omnivggt, written to data/rmvd/eth3d). If the download stops, rerun the script and it resumes:

hf download tarek199147/lens3d scripts/download_rmvd.sh --repo-type dataset --local-dir .
bash scripts/download_rmvd.sh

Examples

Intermediate results for the same view (courtyard/DSC_0286). The evidence is the same at both ratios; only the selected patches and the masked input change between 10% and 25%.

Edges & Corners

	Corners	Edges	Score
Evidence
	Selected patches	Masked input
10%
25%

Geometric Correspondences

	Keypoints	Matches
Evidence
	Selected patches	Masked input
10%
25%

Random (Low-evidence)

	Protected local	Protected correspondences	Eligible patches
Evidence
	Selected patches	Masked input
10%
25%

Data layout

data/rmvd/
├── eth3d/                          original ETH3D (scripts/download_rmvd.sh)
├── eth3d_local_features_r10/
│   ├── config.json                 generation parameters and summary
│   ├── selections.json             per-image masking statistics
│   ├── masks/<scene>/images/dslr_images/DSC_XXXX.png   binary mask, original resolution (255 = masked)
│   └── vis/                        visualizations
├── eth3d_local_features_r25/
├── eth3d_correspondences_r10/
├── eth3d_correspondences_r25/
├── eth3d_random_r10/
└── eth3d_random_r25/

Download

Run from the Lens3D root, so the masked datasets are placed in data/rmvd/ next to eth3d/:

hf download tarek199147/lens3d --repo-type dataset --include "eth3d_*" --local-dir data/rmvd

Reproducibility

The complete implementation, environment setup, masking generation, and evaluation pipeline are available in the Lens3D GitHub repository.

4. Results and Analysis

After the data is generated, we proceed to evaluate the models on the masked datasets and analyze the impact from MV-Reconstruction and Camera Pose estimation, following the evaluation protocol described in Amb3r (AMB3R: Accurate Feed-forward Metric-scale 3D Reconstruction with Backend).

4.1 Multi-view 3D Reconstruction

We report the Absolute Relative Error (Abs), reconstruction Accuracy (Acc), and Completeness (Comp) as measures of reconstruction error.

Lower values indicate better performance.

10% Masking

Model	Clean			Edges & Corners			Geometric Correspondences			Random (Low-evidence)
	Abs	Acc	Comp	Abs	Acc	Comp	Abs	Acc	Comp	Abs	Acc	Comp
VGGT-1B	6.17	0.1284	0.1196	6.24	0.1359	0.1305	5.75	0.1317	0.1222	6.40	0.1292	0.1287
AMB3R	4.63	0.0998	0.0973	5.23	0.1091	0.1060	5.01	0.1054	0.1021	4.88	0.1029	0.1002
DA3	9.23	0.2118	0.1128	9.65	0.2204	0.1183	8.85	0.2031	0.1056	9.41	0.2168	0.1130
VGGT-Omega	3.56	0.0726	0.0747	3.82	0.0762	0.0757	3.72	0.0755	0.0749	3.75	0.0760	0.0763

Abs	Acc
Comp	Sensitivity

Multi-view 3D reconstruction under 10% masking

25% Masking

Model	Clean			Edges & Corners			Geometric Correspondences			Random (Low-evidence)
	Abs	Acc	Comp	Abs	Acc	Comp	Abs	Acc	Comp	Abs	Acc	Comp
VGGT-1B	6.17	0.1284	0.1196	9.15	0.1682	0.1731	6.72	0.1455	0.1499	6.14	0.1367	0.1313
AMB3R	4.63	0.0998	0.0973	7.02	0.1347	0.1330	6.02	0.1206	0.1166	5.43	0.1111	0.1075
DA3	9.23	0.2118	0.1128	11.10	0.2410	0.1286	9.79	0.2231	0.1147	9.43	0.2230	0.1123
VGGT-Omega	3.56	0.0726	0.0747	4.70	0.0927	0.0879	4.12	0.0865	0.0813	4.01	0.0826	0.0808

Abs	Acc
Comp	Sensitivity

Multi-view 3D reconstruction under 25% masking

4.2 Camera Pose Estimation

Similarly, we evaluate camera pose estimation using Relative Rotation Accuracy (RRA@5), Relative Translation Accuracy (RTA@5), and mean Average Accuracy (mAA@30).

Higher values indicate better performance.

10% Masking

Model	Clean			Edges & Corners			Geometric Correspondences			Random (Low-evidence)
	RRA@5°	RTA@5°	mAA@30°	RRA@5°	RTA@5°	mAA@30°	RRA@5°	RTA@5°	mAA@30°	RRA@5°	RTA@5°	mAA@30°
VGGT-1B	95.94	80.63	86.55	95.38	79.41	86.01	95.21	79.13	86.15	95.44	79.90	86.20
AMB3R	97.29	81.96	87.62	96.47	80.45	86.68	96.40	80.56	87.00	96.98	80.86	87.23
DA3	91.85	83.01	86.60	91.89	82.41	85.84	91.82	83.27	86.89	91.50	82.27	86.17
VGGT-Omega	98.60	90.96	92.71	98.86	90.12	92.29	99.04	90.84	92.70	98.48	89.88	92.35

RRA@5°	RTA@5°
mAA@30°	Sensitivity

Camera pose estimation under 10% masking

25% Masking

Model	Clean			Edges & Corners			Geometric Correspondences			Random (Low-evidence)
	RRA@5°	RTA@5°	mAA@30°	RRA@5°	RTA@5°	mAA@30°	RRA@5°	RTA@5°	mAA@30°	RRA@5°	RTA@5°	mAA@30°
VGGT-1B	95.94	80.63	86.55	92.57	72.87	82.78	93.37	75.07	84.62	94.09	78.36	85.62
AMB3R	97.29	81.96	87.62	94.79	75.26	84.45	95.72	77.78	85.59	95.30	78.36	86.39
DA3	91.85	83.01	86.60	91.43	81.42	84.99	90.72	81.38	85.69	91.85	81.71	86.03
VGGT-Omega	98.60	90.96	92.71	98.50	87.74	91.27	99.14	88.95	92.13	98.29	88.50	91.67

RRA@5°	RTA@5°
mAA@30°	Sensitivity

Camera pose estimation under 25% masking

4.3 Key Findings and Discussion

Evidence sensitivity. We measure degradation from each model's baseline (clean condition) using the harmonic mean of Accuracy and Completeness for reconstruction, and mAA@30° for camera pose.

Higher positive percentages indicate stronger impact from removing the targeted visual evidence.

10% masking	25% masking

Evidence sensitivity of multi-view reconstruction (top) and camera pose estimation (bottom) under 10% and 25% masking

Our experiments, in general, show that feed-forward 3D reconstruction models are highly sensitive to the removal of distinctive local image structures and correspondence-rich regions, while low-evidence regions have a comparatively smaller impact.

• Edges and corners have the strongest overall impact, followed by geometric correspondences. At 25% masking, reconstruction accuracy error (Acc) increases by 13.8–35.0% when removing edges and corners, compared with 5.3–20.8% for correspondence-rich regions across VGGT, AMB3R, DA3, and VGGT-Omega. This suggests that distinctive local structures (edges and corners) are particularly important for feed-forward 3D reconstruction, while geometric correspondence evidence also contributes significantly.

• Low-evidence regions have the least impact. At 25% masking, removing randomly selected low-evidence regions increases Acc error by 5.3–13.8%, generally less than evidence-targeted masking. This suggests that reconstruction performance depends not only on the amount of missing information but also on the visual evidence contained in the removed regions.

• Missing visual evidence appears partially recoverable. At 10% masking, most models show a modest degradation in reconstruction and pose accuracy, suggesting they can infer missing information from the remaining evidence, possibly using learned priors (inpainting).

5. Conclusion

What can we conclude? Feed-forward 3D reconstruction models are generally more sensitive to removing edges and corners than correspondence-rich or low-evidence regions. This suggests that distinctive local structures play an important role in their geometric predictions, potentially supporting correspondence-like reasoning. Their relative robustness at 10% masking indicates that the remaining visual evidence can often compensate for partial information loss.

What remains unclear? Our experiments measure output-level sensitivity without inspecting latent representations or token-level interactions. We therefore cannot determine whether models implicitly establish correspondences or encode epipolar constraints, nor how robust these internal representations are to missing evidence.

Further validation should include DTU and 7-Scenes for reconstruction and longer sequences such as RealEstate10K for camera pose estimation. More diverse scenes, controlled masking distributions, and uncertainty estimates are also needed to establish the generality of these findings.

6. Future Research Directions

6.1 Lens3D: Further Investigations

We aim to continue Lens3D by investigating how feed-forward 3D reconstruction models exploit geometric evidence and how their internal representations evolve, focusing on three directions:

• Multi-view geometric understanding and robustness: Evaluate challenging real-world and synthetic scenes with repetitive textures, ambiguous correspondences, and difficult camera configurations. Systematically vary image ordering, view overlap, and parallax to identify geometric failure modes and assess reconstruction consistency.

• Internal geometric representations: Inspect latent tokens and cross-view attention under clean and masked inputs, using correspondence probes to understand how geometric relationships are encoded and affected by missing evidence.

• Layer-wise geometric reasoning: Track correspondence quality and geometric consistency across network layers to determine whether successive operations progressively refine geometry or become redundant, and whether larger models provide meaningful improvements.

These investigations will help us better understand the strengths and limitations of new feed-forward 3D models, identify ways to improve their efficiency and 3D representations, and explore how they can be integrated into existing pipelines such as SLAM and SfM.

6.2 Geometry-Aware Fine-Tuning

Motivated by our findings, we propose geometry-aware fine-tuning to improve reconstruction robustness when important visual evidence is missing. Instead of conventional random masking, we selectively remove regions containing edges, corners, and geometric correspondences in some views, while keeping them visible in overlapping views. Using clean 3D geometry and camera poses as supervision, the model learns to reconstruct masked regions by exploiting complementary geometric evidence from other views. This extends the idea of masked reconstruction to targeted multi-view geometric supervision.

The goal is to encourage cross-view geometric reasoning rather than relying heavily on distinctive features within individual images. Inspired by EPO: Boosting 3D Foundation Models with Edge-based Pose Optimization, which demonstrates the value of edge-based constraints for pose refinement, we also propose incorporating edge-based geometric consistency losses to encourage accurate alignment of predicted structure across views.

We would evaluate this approach by comparing reconstruction accuracy, camera pose estimation, and sensitivity to missing evidence before and after fine-tuning, while ensuring that performance on clean images is preserved.