PaperlayerView exact revision on arXiv

Source-linked AI summary

VGGT-$\Omega$

Jianyuan Wang, Minghao Chen, Shangzhan Zhang, Nikita Karaev, Johannes Schönberger, Patrick Labatut, Piotr Bojanowski, David Novotny, Andrea Vedaldi, Christian Rupprecht

arXiv:2605.15195v1cs.CV

TL;DR

Feed-forward reconstruction models are competitive, but how their quality scales with model and data size remains less explored in 3D vision. VGGT-Ω combines architectural changes, large-scale annotations, and self-supervised training; it achieves state-of-the-art results across static and dynamic benchmarks, while its registers also support downstream uses.

  • Problem

    The paper asks how feed-forward reconstruction quality scales with model and data size, a question less explored in 3D vision than in foundation models.

    Source: §1 · p. 1

  • Method

    VGGT-Ω combines register attention and streamlined prediction heads with large-scale annotated data and teacher-student self-supervised training.

    Source: §1 · p. 2, §7 · p. 15

  • Results

    VGGT-Ω achieves state-of-the-art results on three static and three dynamic benchmarks; on Sintel, camera-estimation AUC@3◦ is 40.0 vs. 22.5.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The learned registers carry useful global information that improves vision-language-action models and supports alignment with language.

    Source: §1 · p. 2, §7 · p. 15

    Self-supervised training has had little impact on most benchmarks so far, and the successful teacher-student approach still requires a pretrained model.

    Source: §5 · pp. 13–14

Abstract

from arXiv · show

Recent feed-forward reconstruction models, such as VGGT, have proven competitive with traditional optimization-based reconstructors while also providing geometry-aware features useful for other tasks. Here, we show that the quality of these models scales predictably with model and data size. We do so by introducing VGGT-$\Omega$, which substantially improves reconstruction accuracy, efficiency, and capabilities for both static and dynamic scenes. To enable training this model at an unprecedented scale, we introduce architectural changes that improve training efficiency, a high-quality data annotation pipeline that supports dynamic scenes, and a self-supervised learning protocol. We simplify VGGT's architecture by using a single dense prediction head with multi-task supervision and removing the expensive high-resolution convolutional layers. We also use registers to aggregate scene information into a compact representation and introduce register attention, which restricts inter-frame information exchange to these registers, in part replacing global attention. In this way, during training, VGGT-$\Omega$ uses only about 30% of the GPU memory of its predecessor, allowing us to train with 15x more supervised data than prior work and to leverage vast amounts of unlabeled video data. VGGT-$\Omega$ achieves strong results for reconstruction of static and dynamic scenes across multiple benchmarks, for example, improving over the previous best camera estimation accuracy on Sintel by 77%. We also show that the learned registers can improve vision-language-action models and support alignment with language, suggesting that reconstruction can be a powerful and scalable proxy task for spatial understanding. Project Page: http://vggt-omega.github.io/

1. Introduction

VGGT-Ω investigates whether feed-forward reconstruction scales with model and data size, using architectural, data, and training changes to support that scaling. It reports state-of-the-art results across static and dynamic benchmarks, alongside useful downstream scene-register features.

1. Introduction

  • The paper asks whether feed-forward reconstruction can scale and what benefits scaling brings, a less-explored question in 3D vision than in foundation models.Reconstruction is framed as a possible proxy task for spatial understanding because learned geometry-aware features have been useful in other tasks.Source: §1 · p. 1
  • Register attention restricts inter-frame exchange to registers, which interact with other tokens within each frame and aggregate scene-level information.Replacing 25% of global-attention layers incurs no measurable performance drop while saving around 23% FLOPs and 16% backbone memory during training; replacing all global-attention layers causes a considerable performance drop.Source: §1 · p. 2
  • Replacing memory-intensive dense-predictor convolutions with one MLP and pixel shuffle reduces memory use without performance degradation.The replaced high-resolution convolutional layers consume disproportionate GPU memory for forward activations despite representing a small share of model parameters.Source: §1 · p. 2
  • VGGT-Ω retains multi-task losses but uses one dense depth head and one sparse camera head instead of redundant dense prediction heads.Depth maps, point maps, and tracking features remain directly supervised through losses.Source: §1 · p. 2
  • The three architectural changes save 70% of GPU memory during training and modestly improve inference speed.The paper also reports that its register design can provide useful features for vision-language-action models and language alignment.Source: §1 · p. 2
  • A high-quality annotation pipeline retains 0.8 million accurately annotated sequences from around 40 million internal Internet-style videos, about one-third dynamic, contributing to four million total annotated scenes or sequences.The combined data is more than 15× the amount used by VGGT.Source: §1 · p. 2
  • As model capacity grows from 0.2B to 10B parameters and training data expands from a few thousand to two million sequences, reconstruction point error improves consistently.Figure 1 averages performance across six datasets; models use approximately the same number of training tokens.Source: Figure 1 · p. 1, §1 · p. 2
  • 77%: On Sintel, camera-estimation AUC@3◦ is 40.0 vs. 22.5, and VGGT-Ω achieves state-of-the-art results across three static and three dynamic benchmarks.It also reports Sintel camera-estimation AUC@30◦ of 79.1 vs. 58.3 and depth-estimation δ1.25 of 93.5 vs. 74.1; it is 50× faster than MegaSaM.Source: §1 · p. 2

2. Related work

Related work traces reconstruction from classical multi-view geometry and SfM to feed-forward systems that jointly process multiple images, and reviews extensions to dynamic scenes and downstream representations. It also covers register tokens and approaches to computational scalability.

2. Related work

  • Classical multi-view geometry led to practical SfM systems, while monocular dynamic reconstruction developed as a distinct line of work addressing changing scene geometry.The passage traces dynamic reconstruction to early work by Bregler et al. and Torresani et al.Source: §2 · pp. 2–3
  • Feed-forward models infer geometry and camera poses directly from images; multi-image systems such as VGGT avoid the post-optimization across views required by earlier image-pair methods.The related work describes VGGT as an early representative that surpassed post-optimization methods using feed-forward inference alone.Source: §2 · p. 3
  • To improve computational scalability, related methods use token merging, sparse attention, descriptor-based aggregation, or bounded state updated over frames.Other approaches process frames causally while caching tokens, key-value pairs, local windows, or maps.Source: §2 · p. 3
  • Feed-forward reconstruction features have been applied beyond geometry to video generation, novel-view synthesis, vision-language and vision-language-action models, and perception tasks.The cited applications include detection, segmentation, matching, occupancy prediction, and place recognition.Source: §2 · p. 3
  • Dynamic reconstruction spans optimization-based methods, feed-forward extensions of pairwise models, and systems designed for incremental or unified 4D reconstruction.Examples include MegaSaM, MonST3R, CUT3R, PI3, DA3, and D4RT.Source: §2 · pp. 3–4
  • Prior ViT studies found that a small number of tokens can encode global information; register tokens separate that role from image tokens and can aggregate information across a sequence.Unlike work that discards registers, this line of work also studies their use in downstream applications.Source: §2 · p. 4

3. Method

VGGT-Ω combines architectural changes that reduce training memory with multi-task supervision, self-supervised learning, and expanded training data. Its pipeline supports reconstruction of both static and dynamic scenes.

3.1.1. Feature Extraction and Tokenization

  • Image tokens from a DINOv3-initialized transformer are combined with camera tokens and scene registers, while register attention exchanges information across frames through registers.Only registers participate in cross-frame self-attention in those blocks; later frame-wise attention redistributes the aggregated information to image tokens.Source: §3.1.2 · pp. 4–5, §3.1.1 · p. 4

3.1. A New Scalable Architecture

  • Global attention attends to all tokens across frames at quadratic cost, and its maps are typically sparse, so the method replaces 25% of global attention layers with register attention restricted to registers across all frames.Source: §3.1.2 · p. 5

3.1.3. Decoding

  • High-resolution DPT convolutional blocks are replaced with an MLP and pixel shuffle, while early low-resolution convolutions remain to avoid blocky depth artifacts in outdoor scenes.The camera head jointly predicts cameras in one pass without iterative refinement.Source: §3.1.3 · pp. 5–6

3.2. Training Losses

  • A single dense depth head retains multi-task supervision of depth, cameras, points, and feature matches, yielding nearly the performance of multiple dense heads with less memory use.The camera loss uses an ℓ1 objective; the depth loss includes uncertainty and gradient-consistency terms.Source: §3.2 · p. 6

3.3. Dynamic Reconstruction

  • For dynamic reconstruction, VGGT-Ω predicts depth and camera parameters rather than explicit motion outputs, avoiding the need for dynamic point maps or motion masks.The paper motivates this choice by noting that scene motion and camera motion can be coupled in representations such as point or ray maps.Source: §3.3 · p. 6

3.4. Self-supervised Training

  • Self-supervised training matches student features and camera and depth predictions to an exponential-moving-average teacher under independent augmentations of the same video frames.Augmentations include color changes, blur, rotations, patch masking, and frame reordering; camera and depth heads are frozen to prevent collapse.Source: §3.4 · p. 6

3.5. Training Data

  • Public and internal datasets provide about 3M sequences, while annotation of roughly 40M Internet-style videos yields about 200K dynamic and 600K static scenes with high-quality camera and depth labels.The annotation pipeline prioritizes quality and discards depth labels in regions likely to be dynamic.Source: §3.5 · p. 7, §3.5.1 · p. 7, §3.5.2 · p. 7

3.5.2. Data Annotation Pipeline

  • Video annotation combines VLM filtering and movable-object masks with feature matching, COLMAP reconstruction, depth consistency checks, and a classifier trained on hand-labeled sequences.The classifier uses camera-up consistency, parallax, and trajectory smoothness features.Source: §3.5.2 · p. 7

4. Experiments

The experiments evaluate VGGT-Ω on static and dynamic reconstruction benchmarks, test its scaling and design choices, and explore registers in downstream tasks. The model performs strongly across benchmarks, while its registers also support language and robotics applications.

4.2. Benchmarking

  • VGGT-Ω outperforms baselines across static and dynamic camera benchmarks; on Sintel, AUC@3◦ rises from 22.5 to 40.0, while δ1.25 improves from 86.1 to 93.5.For depth on Sintel, AbsRel falls from 0.118 to 0.081; the reported camera-estimation AUC@3◦ gain is 77%.Source: Table 1 · p. 8, §4.2 · pp. 8–9

4.1. Implementation Details

  • Training used 200M, 500M, 1B, and 10B variants, with supervised and self-supervised stages, on 128 96GB H100 GPUs.Optimization ran for 240K iterations, with 160K supervised, 50K self-supervised, and a final 30K supervised stage.Source: §4.1 · p. 8

4.2. Benchmarking

  • VGGT and VGGT-Ω handle around 1,250 frames on one A100, versus about 750 for DA3; VGGT-Ω is faster, but replacing all global attention lowers reconstruction accuracy.Replacing 25% of global attention layers with register attention yields a reported 20–25% speedup; the all-register variant reduces 1,000-frame runtime from 240.2 to 11.7 seconds.Source: §4.2 · p. 9, Figure 7 · p. 10

4.3. Ablation Studies

  • Increasing training sequences in 10× steps lowers point error from 0.275 to 0.073, consistent with power-law-like improvement as model or data size grows.Ablations also report point error of 0.071 for global attention only and 0.073 with 25% register attention; removing point and matching losses raises it from 0.073 to 0.078, while self-supervised training reduces it to 0.070.Source: §4.3 · p. 10
  • On Sintel, the annotation pipeline scores 96.4% AUC@30◦ for camera pose and 99.3% δ1.25 for depth, versus MegaSaM’s 62.1% and 77.2%.The pipeline excludes ambiguous sequences or pixels rather than maximizing annotation yield.Source: §4.3 · pp. 10–11

4.4. Applications of Registers

  • Geometry-aware registers improve OpenVLA-OFT success rates across all LIBERO tasks when added as scene tokens.The model is frozen in this experiment; Table 3 reports LIBERO success rates.Source: §4.4 · p. 11, Table 3 · p. 11
  • After language alignment, register-derived embeddings reach 76.8% top-1 and 97.0% top-3 retrieval accuracy, with a text-only LLM embedding yielding 47.5% and 77.8% without additional training.The language token reads only registers, and alignment fine-tuning does not degrade geometric tasks.Source: §4.4 · pp. 11–12
  • Registers can also support sequence- or frame-level prediction through task-specific tokens; preliminary tests cover metric scale, gravity direction, and human presence.These applications are presented as promising possibilities for further exploration.Source: §4.4 · p. 12

5. Further Insights

Further analyses probe where VGGT-Ω stores information, how its representations track motion, and how design and training choices affect efficiency and behavior. The section also reports practical findings and limits for self-supervised training and dense prediction.

Model Souping: Where Is the Information Stored?

  • A 50%-50% fusion of VGGT frame-wise attention-block FFNs resolves certain noisy depth errors, while Q/K/V fusion does not; camera extrinsics appear encoded more broadly.Depth and field-of-view information is largely stored in FFNs, and variable-frame generalization is also linked to frame-wise attention blocks.Source: §5 · p. 12

Motion Awareness

  • PCA clustering tracks the moving dancer without labels, optical flow, learned probes, or temporal-order input, while deeper layers shift toward highlighting all people.Early layers isolate the dancer most cleanly; middle layers retain a weaker motion signal, and deepest layers yield more global, semantic responses.Source: §5 · pp. 12–13

Normalizing Predictions

  • Prediction normalization leaves converged quantitative performance unchanged but produces more spatially spread point clouds and less stable optimization.The authors suggest pretrained initialization when normalization is needed.Source: §5 · p. 13

DINO Initialization; Invalid Area Prediction

  • DINOv3 initialization eases optimization, with training from scratch requiring 4–8× more iterations; an invalid-region branch predicted masks accurately but did not remove sky from foreground depth.The authors therefore omit the invalid-region predictor from the final model.Source: §5 · p. 13

Register Attention Only; Auxiliary Inputs

  • Register attention alone cuts FLOPs to 6% of the original model but reduces performance to original VGGT levels; auxiliary inputs helped task-specific fine-tuning but often harmed pretraining.The authors describe register-only attention as a possible on-device trade-off and auxiliary-input results as preliminary.Source: §5 · p. 13

Synthetic and Real Data Mixture; How to Fine-Tune VGGT/VGGT-Ω

  • Synthetic data contributes more directly to accuracy and real data to generalization; the authors recommend roughly 80% synthetic and 20% real data per epoch.For reconstruction fine-tuning, they recommend a full warm-up-and-cosine schedule and initially removing the uncertainty loss, especially for very small datasets.Source: §5 · p. 13

Self-Supervised Training

  • Student–teacher self-supervision improved generalization, especially out of distribution, but had little effect on most benchmarks and still requires a pretrained model.Other explored protocols did not help in the authors’ implementation; self-supervised reconstruction remains open.Source: §5 · pp. 13–14

Dense Prediction Heads

  • MLP-only dense heads produced visible depth-map block artifacts, especially outdoors with distant objects, despite often better metrics, speed, memory use, and gradient stability.The chosen compromise retains shallow low-resolution convolutions, which improved spatial smoothness; MLP-only heads remain a direction for exploration.Source: §5 · p. 14

6. Discussion

The discussion weighs simplicity against task-specific improvements and describes feed-forward reconstruction’s speed, robustness, and representational benefits. It also considers optimization as a complement and unified multimodal models as a possible direction for reconstruction.

Prioritizing Simplicity

  • Simpler VGGT-Ω heads sacrifice some task-specific gains, but a well-trained backbone allows new or improved heads to be trained in 5–10K iterations.Iterative camera refinement and raw-RGB injection into the depth head improved AUC@3° by 4%–6% and δ1.25 by about 2% across multiple datasets.Source: §6 · p. 14

Benefits of Feed-forward Reconstruction

  • Feed-forward reconstruction is tens to hundreds of times faster in some settings, handles little or no parallax and dynamic scenes, and provides geometry-aware representations for 3D tasks.The authors present such reconstruction as a promising proxy task for general-purpose 3D representations.Source: §6 · p. 14
  • Feed-forward reconstruction can complement optimization: bundle adjustment can reach angular errors of a few hundredths of a degree in well-conditioned settings, while feed-forward methods can initialize later optimization.The discussion cites NeRFs and Gaussian splatting as applications where that precision remains valuable.Source: §6 · pp. 14–15

The Role of 3D/4D in the Era of Large Models

  • The authors envision unified multimodal models combining reconstruction with other capabilities, potentially drawing on text and video data and cross-task consistency.They suggest reconstruction-oriented data could be added without major architectural or training-framework changes, while framing the direction as prospective.Source: §6 · p. 15

7. Conclusion

The conclusion presents VGGT-Ω as a scaled feed-forward reconstruction model for static and dynamic scenes. It summarizes its architectural, data, and training changes, reports predictable scaling, and points to learned registers as useful beyond geometry.

7. Conclusion

  • VGGT-Ω combines register attention, a single multi-task-supervised dense head, dynamic-scene annotation, and self-supervised training on unlabeled video.The authors report that these changes enabled training at unprecedented scale.Source: §7 · p. 15
  • VGGT-Ω achieves strong results across static and dynamic benchmarks, and its quality scales predictably with model capacity and data size.Source: §7 · p. 15
  • Learned registers carry useful global information, improving vision-language-action models and supporting alignment with language.Source: §7 · p. 15

Supplementary Material

The supplement points readers to implementation details, data issues, and limitations. Its additional-details section enumerates training, pairing, prompting, filtering, and language-alignment procedures.

Supplementary Material

A. Additional Details

  • Additional details address training setup and architecture, positive and negative pairs, the VLM prompt, annotation filtering, and language alignment.Source: A. Additional Details · p. 24

A.1. Training and Architecture

The training setup combines ground-truth normalization, loss weighting, and stabilization with variable-frame sampling and data augmentation. Video frames are sampled locally in time, including challenging sequences with little or no visual overlap.

A.1. Training and Architecture

  • Video sequences are sampled from a local temporal window, and the authors report that challenging subsets with little or no visual overlap improve generalization.The number of input frames is sampled from 1 to 24; batch size is set to saturate GPU memory.Source: §A.1 · p. 24
  • Ground-truth quantities are transformed into the first camera’s frame and scaled to unit space, while predictions are not normalized.The scale is the average distance of ground-truth 3D points to the origin; depth maps and translation vectors are scaled.Source: §A.1 · p. 24
  • Training uses loss weights of λcam = 5.0, λdepth = 1.0, λpoint = 0.5, and λmatch = 0.1, with gradient clipping at 1.0 and QKNorm in attention.Source: §A.1 · p. 24
  • Frames are independently masked with probability 0.05, and color jitter uses brightness, contrast, and saturation of 0.5 and hue of 0.1.Masked regions are blacked out and their corresponding depth values marked invalid.Source: §A.1 · p. 24
  • During self-supervised training, the teacher is updated by exponential moving average with decay 0.999; the camera head uses ReLU for focal length and no activation for quaternion or translation.Source: §A.1 · p. 24

A.2. Constructing Positive and Negative Pairs

Positive token pairs are formed from depth-unprojected points that project consistently across views, while negatives must satisfy both geometric and appearance constraints. A VLM prompt screens videos for reconstruction suitability and labels scene dynamics.

A.2. Constructing Positive and Negative Pairs

  • Positive pairs link query and target patches when depth-consistent projections yield more than 10% overlap.Valid query pixels are assigned to patches, projected into other frames using camera parameters, and filtered by image bounds, depth consistency within 1%, and a 4-pixel boundary exclusion; eligible query patches are sampled in proportion to correspondence counts.Source: §A.2 · p. 24
  • Negative pairs are sampled to satisfy both a large epipolar/Sampson distance and a sufficiently large difference in mean RGB, then balanced with positives for matching-loss supervision.The supervision applies to last-layer tokens.Source: §A.2 · p. 24

A.3. VLM Prompt

  • The VLM prompt assigns ACCEPT, REJECT_HARD, or REJECT_SOFT and records whether the scene is static or dynamic.Hard rejection covers editing, nonphysical content, severe visual failures, major obstructions, and uncalibrated non-pinhole footage; soft rejection covers weak parallax, texture issues, reflections or transparency, focus problems, and dynamic dominance.Source: §A.3 · pp. 24–25

A.4. Annotation Filtering Criteria

The annotation-filtering pipeline extracts geometric signals from reconstructions to identify common failure modes. Its features assess trajectory behavior, parallax, point-cloud shape and noise, and dense-depth completeness; its depth metrics assume dynamic pixels are masked.

A.4. Annotation Filtering Criteria

  • High trajectory acceleration scores indicate jittery camera paths often associated with poor SfM convergence.Translation smoothness is the mean squared second-order difference of translation vectors; rotation smoothness is computed similarly from rotation vectors.Source: §A.4 · p. 25
  • Low median maximum parallax across visible camera pairs indicates rotation-only motion or extreme distance, both problematic for multi-view stereo.Source: §A.4 · p. 25
  • High point-cloud linearity is flagged as a degeneracy associated with fly-by reconstructions and cylindrical ambiguity.PCA features include linearity (v1 − v2)/v1, planarity (v2 − v3)/v1, and scattering v3/v1.Source: §A.4 · p. 25
  • Low average valid-depth completeness can signal multi-view stereo failure, often due to textureless surfaces, specularities, or bad cameras.Source: §A.4 · p. 25
  • The point-cloud noise feature is the percentage of points whose average nearest-neighbor distance exceeds the global average by more than two standard deviations.The pipeline’s depth values cover rigid pixels only, so its metrics assume dynamic pixels are already masked; depth for dynamic pixels is assumed to be learnable from synthetic data.Source: §A.4 · p. 26

A.5. Language Alignment

Language alignment uses a fixed VLM to produce a sequence-level language embedding, which is mean-pooled from generated text-token hidden states and projected. The prompt asks for a concise, coherent description of shared scene content and stable appearance across views.

A.5. Language Alignment

  • A fixed VLM generates up to 128 tokens from an unordered image collection, and mean-pooled generated-text hidden states are projected into a sequence-level language embedding.Prompt and image tokens are excluded from pooling; the VGGT-Ω side is optimized during alignment fine-tuning.Source: §A.5 · p. 26
  • The prompt requests one short sentence each for scene, content, and appearance, emphasizing shared multi-view information rather than frame-specific changes.It excludes camera motion, viewpoint order, image quality, and uncertainty, and mentions dynamic objects only when prominent and identifiable across views.Source: §A.5 · p. 26
  • Representative outputs describe a red-lit patterned room, a snowy forest and winding road, a tiled outdoor picnic area, and an urban plaza with a fountain.Source: §A.5 · p. 26

B. Data Quality

The section documents how noisy or inconsistent training annotations create characteristic reconstruction failures, some of which may not appear in standard benchmark results.

Data Quality

  • Shot transitions in self-supervised videos can trigger sharp loss spikes and may cause exploding gradients, so training is restricted to VLM-prefiltered videos.The pre-filter reliably removes such difficult cases.Source: B. Data Quality · p. 26
  • Annotation noise can be memorized and cause failures on new samples resembling mislabeled training images, even when standard benchmarks do not reveal them.The model may learn general reconstruction principles while also memorizing idiosyncratic noise.Source: B. Data Quality · pp. 26–27

Sensors

  • Sensor capture errors can create foreground-background leakage and spurious depth structures, leading the authors to exclude such datasets in later training.The figure identifies ScanNet++ as the top-row example.Source: B. Data Quality · p. 27, Figure 10 · p. 27

Thin Structure

  • Synthetic depth labels for thin objects can be incomplete, smoothed, or misaligned, and models may ignore or wash out those structures.The authors hypothesize that higher-resolution inputs or better synthetic data could mitigate this problem.Source: B. Data Quality · p. 27, Figure 10 · p. 27

Fake Background

  • Synthetic background depth may encode proxy dome or floor geometry rather than scene structure; foreground thresholding filters clear boundaries, and remaining cases are excluded.The figure identifies the doming effect among its examples.Source: B. Data Quality · p. 27, Figure 10 · p. 27

Doming Effect

  • Weak global constraints can let bundle adjustment produce locally plausible but globally curved geometry, which is harmful as supervision and can be filtered.The section cites near-parallel views, insufficient loop closure, small triangulation angles, and camera self-calibration errors as contributing conditions.Source: B. Data Quality · p. 27, Figure 10 · p. 27

Humans in Walls

  • Human-boundary depth errors can make models absorb pedestrians into nearby buildings; excluding or re-estimating MegaDepth reduces this artifact.The described mismatch can arise when patch matching assigns boundary pixels to static architecture.Source: B. Data Quality · p. 27

Data Ambiguities

  • Synthetic window annotations may label the window surface while other datasets label objects visible through it, creating inconsistent depth supervision.The section reports that this mismatch occasionally leads to strange predictions.Source: B. Data Quality · p. 27

C. Limitations

The authors report limitations involving challenging visual conditions, training-data distribution, and masked sensitive content.

C. Limitations

  • Strong motion blur, abrupt field-of-view changes, and highly distorted cameras are settings where reconstruction quality drops; noisy early training data is linked to instability in monitor-heavy offices.The authors attribute these limitations primarily to the training-data distribution and hope to address them with more diverse, challenging sequences.Source: C. Limitations · p. 28
  • Privacy- or licensing-driven masking of faces and trademarks can lead to rare artifacts or unstable predictions in masked regions.The authors note that predicted depth for a person in black clothing can occasionally become less smooth, producing visually distorted outputs.Source: C. Limitations · p. 28
Loading 2605.15195v1…