acceptodds
Under review as a conference paper at ICLR 2027

Can 3D Geometry-aware Features Help with Affordance Prediction?

Abstract

Visual affordance prediction enables embodied agents to identify actionable objects and localize interaction regions from limited visual observations. However, existing affordance prediction methods typically rely on RGB appearance cues or explicit 3D representations. Since embodied agents often need to act before a complete 3D scene model is available, extracting useful geometric information directly from RGB observations is particularly valuable. To this end, we integrate reconstruction-pretrained geometry-aware features with visual representations without constructing an explicit 3D scene model. We evaluate this integration across multiple datasets, predictors, and prediction settings. Our experiments indicate that geometry-aware features can improve visual affordance prediction. Furthermore, the gains depend on data characteristics and view sampling and do not increase monotonically with the number of views. Notably, fine-grained geometric features provide greater benefits than global geometric features.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
visual affordance prediction, geometry-aware representation, 3D, representation learning, computer vision

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_B9miOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether reconstruction-pretrained geometry-aware features can improve visual affordance prediction. The proposed approach uses a frozen VGGT-Ω encoder to extract intermediate features from a target image and, when available, auxiliary views. These features are projected through lightweight adapters and fused with existing visual representations used by several affordance predictors. The method is evaluated on 3DOI, HANDAL, and AGD20K, covering scene-level images/videos, object-centric affordance segmentation, and weakly supervised affordance grounding. The paper studies single-view versus multi-view inputs, different VGGT-Ω feature depths, and several fusion strategies. The main findings are that VGGT-Ω features generally improve affordance prediction, additional views do not always monotonically improve performance, and shallow/multi-layer features can be more useful than deeper individual layers.

Critical strengths

The paper is clearly written and evaluates the proposed feature augmentation across several datasets and existing affordance predictors rather than relying on a single benchmark. The experiments cover different input regimes, including single-view and multi-view settings, and include analyses of feature depth and fusion strategy. The paper also makes a reasonable effort to keep the pretrained encoders frozen and isolate the effect of lightweight feature adaptation, which makes the empirical comparisons relatively easy to understand.

Critical weaknesses
  1. The main research question has already been widely studied, so the conceptual novelty is limited. The paper asks whether 3D geometry-aware features can improve affordance prediction. However, prior work has already shown that geometric/shape cues [1], depth-based geometric priors [2], and geometry encoded in frozen visual foundation models [3] can improve affordance prediction. More broadly, the geometric awareness of foundation-model features has also been systematically studied [4], and VGGT has already demonstrated transferability to downstream tasks [5]. Thus, the main difference here appears to be using VGGT-Ω intermediate features as the geometry source and fusing them into existing affordance predictors. This feels more like a feature/backbone substitution than a new methodology or scientific insight.
  2. Some key experimental protocols are not sufficiently convincing or clearly specified. The HANDAL evaluation is particularly concerning. Table 4 reports AffordanceNet at 60.3/60.8 under the original setting, but only 31.0/13.3 under the authors' split, after which VGGT-Ω improves the result to 67.5/69.7. Such a large baseline drop makes it difficult to interpret the magnitude of the reported improvement without a clearer explanation of how the new split differs from the standard evaluation.
    Moreover, the HANDAL sampling protocol is described inconsistently: Appendix A.3 uses auxiliary offsets of ±4, ±8, and ±12, while Appendix B.2 describes non-overlapping 11-frame segments with the sixth frame as query and stride-based sampling. The paper also does not report repeated runs or variance, making it hard to judge the robustness of the smaller improvements.

[1] Affordance Detection of Tool Parts from Geometric Features, ICRA 2015.

[2] Learning Precise Affordances from Egocentric Videos for Robotic Manipulation, ICCV 2025.

[3] Probing and Bridging Geometry-Interaction Cues for Affordance Reasoning in Vision Foundation Models, CVPR 2026.

[4] Probing the 3D Awareness of Visual Foundation Models, CVPR 2024.

[5] VGGT: Visual Geometry Grounded Transformer, CVPR 2025.

Questions for authors

What is the main scientific novelty beyond showing that VGGT-Ω intermediate features also transfer to affordance prediction? How does this differ from prior work using geometric/depth cues or geometry encoded in frozen foundation models?

In Table 4, AffordanceNet drops from 60.3/60.8 to 31.0/13.3 under “our split,” while adding VGGT-Ω raises it to 67.5/69.7. Why does the baseline deteriorate so substantially under the new split, and how much of the reported gain depends on this evaluation protocol?

Please clarify the HANDAL auxiliary-view sampling protocol. Appendix A.3 uses offsets of ±4, ±8, and ±12, while Appendix B.2 describes non-overlapping 11-frame segments with the sixth frame as query and stride-based sampling. How do these two descriptions correspond?

Were the reported results obtained from a single run? If so, could the authors report results across multiple seeds or provide variance, especially for the relatively small gains in some of the multi-view and fusion comparisons?

5 comments
AuthorsOpenReview ↗

Response to Reviewer B9mi’s Question 1

Q1: What is the main scientific novelty beyond showing that VGGT-Ω intermediate features also transfer to affordance prediction? How does this differ from prior work using geometric/depth cues or geometry encoded in frozen foundation models?

We thank the reviewer for highlighting the contributions of related work. Our contribution is not a new reconstruction model or fusion operator. Rather, we investigate when geometry-aware features pretrained for reconstruction help affordance prediction and how to use them without explicitly reconstructing a 3D scene. Across diverse datasets and models with substantially different architectures, we examine single-view inputs versus related views from the same scene, feature levels, and fusion strategies to understand the role of 3D geometric information in different types of data.

Whereas [1]–[3] use local shape, estimated depth, or geometric cues from frozen models, our work focuses on the conditions under which reconstruction-pretrained geometric features are useful for affordance prediction. Works [4] and [5] do not examine this behavior in affordance prediction. The table summarizes the distinctions between our work and the studies you mentioned. We will state our contribution more clearly as empirical findings and an analysis of the underlying mechanisms, rather than as a new fusion algorithm.

WorkGeometry sourceUse of related viewsCentral research question
Myers et al. (2015) [1]Local shape features from RGB-D dataNot usedCan tool-part geometry support affordance detection?
GAT (Li et al., 2025) [2]Image-estimated depthNot usedCan depth priors improve fine-grained affordance prediction?
Zhang et al. (2026) [3]Geometric cues within frozen foundation modelsNot usedHow can geometric and interaction cues be combined for affordance reasoning?
El Banani et al. (2024) [4]Geometric cues within visual foundation modelsCross-view analysisHow much 3D information do visual foundation models encode?
VGGT (Wang et al., 2025) [5]Representations learned through 3D reconstruction pretrainingMulti-view inputsCan VGGT representations transfer to downstream tasks?
OursIntermediate VGGT-Ω featuresSingle-view inputs and related views of the same sceneWhen do 3D geometric features improve affordance tasks?

[1] Affordance Detection of Tool Parts from Geometric Features, ICRA 2015.

[2] Learning Precise Affordances from Egocentric Videos for Robotic Manipulation, ICCV 2025.

[3] Probing and Bridging Geometry-Interaction Cues for Affordance Reasoning in Vision Foundation Models, CVPR 2026.

[4] Probing the 3D Awareness of Visual Foundation Models, CVPR 2024.

[5] VGGT: Visual Geometry Grounded Transformer, CVPR 2025.

AuthorsOpenReview ↗

Response to Reviewer B9mi’s Question 2

Q2:In Table 4, AffordanceNet drops from 60.3/60.8 to 31.0/13.3 under “our split,” while adding VGGT-Ω raises it to 67.5/69.7. Why does the baseline deteriorate so substantially under the new split, and how much of the reported gain depends on this evaluation protocol?

We identified the cause of the anomalous results. The low reproduced scores of 31.0/13.3 were primarily due to an incompatible inference environment. To accommodate VGGT-Ω, we used a different version of Transformers from the one specified by the official AffordanceNet implementation. This changed CLIP’s forward pass and attention behavior; numerical differences under bf16 further affected the [AFF] features and mask decoding. As a result, the model incorrectly labeled desks, floors, walls, and nearby objects as foreground, producing many false positives. We have replaced Table 4 with the corrected evaluation results.

The performance gains persist under both evaluation protocols. After resolving the environment issue, we reran the experiments using Our Split. We also reproduced the original AffordanceNet sampling procedure as closely as possible for Original Split: within each complete video, we selected auxiliary frames at offsets of ±4, ±8, and ±12 from the target frame tt. We excluded only target frames near video boundaries for which the full set of auxiliary frames was unavailable, and did not subsample the target frames at fixed intervals. The improvement from adding VGGT-Ω remained under Original Split.

MethodsViewgIoU (Our Split) ↑gIoU (Original Split) ↑cIoU (Our Split) ↑cIoU (Original Split) ↑
AffordanceNetSingle60.0159.3559.9858.95
+ Geometry FeatureSingle67.5069.6669.7072.08
+ Geometry Feature1+268.1070.6771.0071.91
AuthorsOpenReview ↗

Response to Reviewer B9mi’s Question 3

Q3: Please clarify the HANDAL auxiliary-view sampling protocol. Appendix A.3 uses offsets of ±4, ±8, and ±12, while Appendix B.2 describes non-overlapping 11-frame segments with the sixth frame as query and stride-based sampling. How do these two descriptions correspond?

We thank the reviewer for pointing out that our description was unclear. The previous sampling procedure had two stages. Appendix B.2 describes target-view sampling: we divided each video into non-overlapping 11-frame segments and selected the sixth frame of each segment as the target view. Appendix A.3 describes auxiliary-view sampling after the target view was selected: we chose frames at offsets of ±4, ±8, and ±12 from the target frame within the complete video. This procedure was intended to minimize repeated use of the same video frames.

We will replace the original experimental results and sampling description with those from the corrected protocol. We appreciate the reviewer’s correction.

Reviewer_9QyDOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper integrates reconstruction-pretrained geometric features with visual representations for affordance prediction, without explicitly reconstructing a 3D scene. Experiments across multiple datasets and settings show that geometric features improve prediction performance, particularly fine-grained features. However, the improvements depend on data characteristics and view sampling, with more views not necessarily leading to better performance.

Critical strengths
  • The paper idea is simple and easy to follow.
  • Overall, I see the idea of incorporating geometric features into affordance learning is interesting. The experiments also provide good justification for the paper.
Critical weaknesses
  • The paper relies on implicit geometric features rather than explicitly modelling 3D geometry. The method relies on VGGT to extract geometric information, which is reasonable but limits the technical novelty of the contribution, as VGGT features is not always guaranteed, raising concerns about the robustness and generalisability.
  • Quite several technical information of the paper is in the Appendix; for example, I can't find the "trainable block" such as 2D Appearance Adaptor or Geometry-aware Adaptor in the main paper.
  • Some design choices seem ad hoc and dataset-dependent (e.g, Equation 4), this raises concerns about the generalization of the method.
  • As the authors mentioned, in several cases, adding more views is not really useful. E.g., Table 3 shows that 1 view already achieves reasonable accuracy, and more views do not significantly improve the results. Is there any justification?
Questions for authors

I have main concerns about the writing, the ad-hoc design based on the dataset, and the results when adding more views.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 9QyD: Weakness 1

W1: The paper relies on implicit geometric features rather than explicitly modelling 3D geometry. The method relies on VGGT to extract geometric information, which is reasonable but limits the technical novelty of the contribution, as VGGT features is not always guaranteed, raising concerns about the robustness and generalisability.

Thank you for raising this constructive concern. We address it in two parts:

-Technical novelty. Our contribution is not a new reconstruction model or fusion operator. Instead, we investigate when 3D or pseudo-3D geometry-aware features pretrained for reconstruction help affordance prediction and how to use them without explicitly reconstructing a 3D scene. We study intermediate features from a frozen VGGT-Ω as a representative case, examining their behavior across datasets, feature levels, and numbers of auxiliary views.

-Effectiveness of VGGT-Ω. VGGT-Ω features reflect learned estimates of 3D structure rather than ground-truth geometry. What they capture varies across samples in focus, detail, and completeness. This variability motivates our analysis of how the features contribute to affordance prediction and yields insights beyond performance gains. For example, adding auxiliary views does not consistently improve performance, particularly in videos containing object interactions. Moreover, a more complete representation of the overall scene does not necessarily improve localization of the interaction-relevant part, and task-relevant local structural details may matter more.

AuthorsOpenReview ↗

Response to Reviewer 9QyD: Weakness 2

W2: Quite several technical information of the paper is in the Appendix; for example, I can't find the "trainable block" such as 2D Appearance Adaptor or Geometry-aware Adaptor in the main paper.

We thank the reviewer for pointing out that the trainable modules were insufficiently described in the main text. We will revise Section 3.3 to specify the inputs, operations, and outputs of both adapter branches, and clarify the fusion interface and trainable modules used for each dataset. The appendix will retain the exact feature dimensions and training configurations.

AuthorsOpenReview ↗

Response to Reviewer 9QyD: Weakness 3

W4: Some design choices seem ad hoc and dataset-dependent (e.g, Equation 4), this raises concerns about the generalization of the method.

We thank the reviewer for pointing this out. Our aim is to study the role of 3D or pseudo-3D geometric features across methods with substantially different architectures. The alternative fusion forms in Eq. (4) are determined by the interfaces of the original baselines; they do not represent different geometric assumptions for different datasets. In all experiments, we use a frozen VGGT-Ω to extract geometric features from the target image and align them with the corresponding visual feature grid.

For 3DOI, the task decoder is trainable, so we concatenate the visual and geometric features and project them to the dimension required by the decoder. For HANDAL, the SAM predictor is tightly coupled to its image embedding and remains frozen. We therefore add the geometric features through a residual projection that preserves the embedding dimension and the original prediction at initialization. AGD20K includes baselines with different feature interfaces, including ResNet, DINOv2, and CLIP, as well as more complex weakly supervised methods. We apply lightweight residual fusion at each baseline’s interface while retaining its original training objective. For these methods, the geometry encoder receives only the target-view image, consistent with their inference settings.

Reviewer_r48oOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates whether intermediate features from a 3D reconstruction model can improve RGB-based affordance prediction without explicitly reconstructing the scene. The authors fuse frozen VGGT-Ω features with existing visual features through adapters. The model takes either a single target image or additional views from the same scene or interaction video, while predicting affordance regions in the target image. Experiments on 3DOI, HANDAL, and AGD20K examine different predictors, view counts, feature depths, and fusion strategies. The results show that geometry-aware features help even with a single image, while adding more views does not consistently improve performance. On HANDAL, shallow features outperform deeper ones, and multi-layer late fusion achieves the best results.

Critical strengths

The research question is clear and practically relevant. The experiments cover multiple datasets and predictors.

Critical weaknesses

The gains are not clearly attributable to geometric knowledge, rather than stronger features or additional trainable capacity. The method mainly extracts multi-layer features from a frozen backbone and integrates them into existing predictors through projection, residual adapters, and late fusion—all established designs. The video experiments mix geometric context with temporal information. Auxiliary frames come from both before and after the target frame. Adding more views also expands the time window, introducing contact, action, and object-state cues. The gains therefore cannot be attributed to complementary geometry alone.

Questions for authors

Are the gains mainly due to geometric knowledge? Could you compare stronger visual features and learnable cross-view attention under matched adapters, feature dimensions, and training budgets? Do multi-frame gains rely on future action cues? Could you compare past-only, future-only, and repeated target frames, and vary frame counts within a fixed time window? Which errors do geometry-aware features reduce? Could you analyze occlusion, geometric ambiguity, and viewpoint changes, and report latency and memory costs alongside performance?

5 comments
AuthorsOpenReview ↗

Response to Reviewer r48o’s Question 1

Q1: Are the gains mainly due to geometric knowledge? Could you compare stronger visual features and learnable cross-view attention under matched adapters, feature dimensions, and training budgets?

We thank the reviewer for suggesting a more rigorous comparison. The additional experiments provide further evidence that the gains primarily stem from geometric knowledge. We compare against stronger visual features and learnable cross-view attention while matching the adapters, fused feature dimensions, and training budgets. For the visual-feature control, we replace the VGGT-Ω branch with DINOv2 features extracted from the same target and auxiliary views, then concatenate the multi-view features. The models use identical adapters, fused feature dimensions, training settings, data splits, and test inputs. Under these matched conditions, adding VGGT-Ω 3D geometry features consistently improves performance across all three subsets. Thus, the gains cannot be explained solely by additional input views, adapter capacity, or training budget.

Taskonomy

MethodsViewsMovable Cat. ↑Localization Box ↑Localization Mask ↑Action Cat. ↑Affordance SIM ↑Affordance Distance ↓Affordance KLD ↓
DINOv2-BSingle View0.85220.60180.60190.92200.12820.05023.5312
DINOv2-LSingle View0.84950.61090.60730.92230.13530.04883.4753
DINOv2-B Concat Views1+20.85400.58740.58690.92440.11990.04863.3891
DINOv2-B Cross-view Attention1+20̲.̲8̲5̲4̲8̲0.59970.61480.92870.11960.04893.3745
DINOv2-B + Geometry FeatureSingle View0.84230̲.̲6̲7̲0̲4̲0̲.̲7̲2̲0̲3̲0̲.̲9̲2̲9̲0̲0.13120.04152.9833
DINOv2-B + Geometry Feature1+20.85890.68200.72780.93220̲.̲1̲3̲3̲2̲0̲.̲0̲4̲1̲8̲3̲.̲0̲0̲1̲1̲

EPIC-Kitchen

MethodsViewsMovable Cat. ↑Localization Box ↑Localization Mask ↑Action Cat. ↑Affordance SIM ↑Affordance Distance ↓Affordance KLD ↓
DINOv2-BSingle View0.87030.62830.61400.97990.04100.07654.3815
DINOv2-LSingle View0̲.̲8̲7̲5̲2̲0.55470.52400.94570.02830.07614.6394
DINOv2-B Concat Views1+20.86930.62610.62560.97590.03540.07404.4746
DINOv2-B cross-view attention1+20.87320.63710.62600.97990.04210.07594.3215
DINOv2-B + Geometry FeatureSingle View0.89220.75800.77840.98390̲.̲0̲4̲5̲6̲0̲.̲0̲6̲4̲4̲4̲.̲0̲4̲8̲3̲
DINOv2-B + Geometry Feature1+20.86930̲.̲6̲8̲7̲1̲0̲.̲7̲6̲6̲4̲0̲.̲9̲8̲0̲9̲0.04700.06304.0325

Articulation

MethodsViewsMovable Cat. ↑Localization Box ↑Localization Mask ↑Action Cat. ↑Affordance SIM ↑Affordance Distance ↓Affordance KLD ↓
DINOv2-BSingle View0.88460.61020.60600.91050.07000.07304.3438
DINOv2-LSingle View0.91790.60270.61850.91440.05650.07754.3820
DINOv2-B Concat Views1+20.92600.61460.61120.91820.05550.07624.5930
DINOv2-B cross-view attention1+20.92010.51560.55560.88470.00260.23956.7276
DINOv2-B + Geometry FeatureSingle View0̲.̲9̲2̲4̲6̲0̲.̲6̲9̲9̲4̲0.74510.93120̲.̲0̲9̲3̲7̲0.06303̲.̲6̲7̲7̲7̲
DINOv2-B + Geometry Feature1+20.91270.70540̲.̲7̲3̲4̲6̲0̲.̲9̲2̲8̲0̲0.10150̲.̲0̲6̲4̲6̲3.5374
AuthorsOpenReview ↗

Response to Reviewer r48o’s Question 2

Q2: Do multi-frame gains rely on future action cues? Could you compare past-only, future-only, and repeated target frames, and vary frame counts within a fixed time window? Which errors do geometry-aware features reduce?

We thank the reviewer for this rigorous question, which helps control the variables more carefully. Following your suggestion, we conducted additional experiments, summarized in the table below. The results show no clear dependence on future frames under the tested settings.

On 3DOI-EpicKitchen, we compared inputs using only past frames, only future frames, or repeated copies of the target frame. We also varied the number of sampled auxiliary frames within a fixed temporal window. We chose this dataset because its videos contain actions and object interactions that change over time. With the trained model held fixed, all three input variants produced results close to those obtained with the original bidirectional frames. Increasing the input from 1+2 to 1+5 frames within the fixed window changed some metrics slightly, but yielded no consistent or substantial benefit from denser sampling.

In particular, repeating the target frame produced similar results. This suggests that the gains in the original multi-frame setting cannot be attributed solely to new action information in the auxiliary frames. One possible explanation is that the sampled frames are close in time to the target frame and therefore have similar visual content, offering limited complementary information. The number of input frames and repeated target representations may also affect feature aggregation.

Past-only, Future-only, and Repeated Target Frames

DatasetViewsModeMovable Cat. ↑Localization Box ↑Localization Mask ↑Action Cat. ↑Affordance SIM ↑Affordance Distance ↓Affordance KLD ↓
Baseline1--0.87030.62830.61400.97990.04100.07654.3815
+ Geometry-aware Feature1+4original0.89740.72830.77130.98150.07060.06053.5569
+ Geometry-aware Feature1+4past0.89540.72810.77070.98150.07020.06103.5643
+ Geometry-aware Feature1+4future0.89740.73020.77230.98050.07050.06163.5575
+ Geometry-aware Feature1+4repeat0.89550.72900.77030.98050.07030.06083.5610

Fixed Time Window

DatasetViewsModeMovable Cat. ↑Localization Box ↑Localization Mask ↑Action Cat. ↑Affordance SIM ↑Affordance Distance ↓Affordance KLD ↓
+ Geometry-aware Feature1+2past 100s0.89440.72750.77050.98150.07020.06143.5625
+ Geometry-aware Feature1+4past 100s0.89540.72810.77070.98150.07020.06103.5643
+ Geometry-aware Feature1+5past 100s0.89640.72910.77100.98150.07010.06123.5645
AuthorsOpenReview ↗

Response to Reviewer r48o’s Question 3

Q3: Which errors do geometry-aware features reduce? Could you analyze occlusion, geometric ambiguity, and viewpoint changes?

We conducted a qualitative analysis and quantitative experiments focusing on the representative cases you identified:

Qualitative analysis. The geometric and multi-view features appear particularly helpful in three cases:

  • Occlusion and blur. Three-dimensional geometry helps distinguish depth ordering and separate the foreground, target object, and background. Auxiliary views may also reveal an occluded target part or show it more clearly. The second and third examples in Figure 5 illustrate improvements under occlusion and blur.
  • Parts that are difficult to distinguish from the whole object. This can occur when the target object is small, the background has complex textures, or the queried part and the rest of the object have similar materials. The first two examples in Figure 4 illustrate improvements associated with the geometric features.
  • Viewpoint changes. Additional views provide complementary information that can help identify the background, localize the target object, and predict its affordance.

Quantitative analysis. For the quantitative analysis, we evaluated samples involving occlusion, geometric ambiguity, and viewpoint changes. We drew the evaluation subset from the original 3DOI-Taskonomy test set, which uses one target view and two auxiliary views. Its static scenes reduce potential confounding factors, while its consistent viewpoint changes facilitate controlled sampling. For each target image and query point, we generated candidates using adjacent views, views with one intervening view, or views with two intervening views, provided that both a preceding and a following real view from the same scene were available. This yielded 4,877 candidates. We then randomly shuffled the test samples and selected, in order, 50 cases with occlusion in the target view and 50 with geometric ambiguity in the target view.

The results show that the most pronounced benefit of adding 3D geometric features is improved localization, particularly on the Mask metric. Affordance prediction also improves consistently. By contrast, the Movable Cat. and Action Cat. classification scores show no consistent gain. Overall, the benefits of geometric features are most reliable for object localization and affordance prediction, whereas their effect on classification varies by condition.

ConditionMethodsMovable Cat. ↑Loc. Box ↑Loc. Mask ↑Action Cat. ↑SIM ↑Peak Distance ↓KLD ↓
OcclusionMulti-view DINOv2-B0.88240.58660.56910.98040.12960.04783.1779
OcclusionBaseline + Geometry Feature0.84310.70010.74010.90200.14100.04222.8441
Geometric ambiguityMulti-view DINOv2-B0.82350.56660.55040.92160.13600.04893.1458
Geometric ambiguityBaseline + Geometry Feature0.86270.61230.69820.92160.14620.04602.8574
Viewpoint change: adjacent viewsMulti-view DINOv2-B0.84640.57460.57240.92860.12200.04863.5079
Viewpoint change: adjacent viewsBaseline + Geometry Feature0.86150.66690.71390.93060.14720.03932.7997
Viewpoint change: skip 1 viewMulti-view DINOv2-B0.86410.57600.56870.92640.12630.04833.4683
Viewpoint change: skip 1 viewBaseline + Geometry Feature0.88040.65860.70870.93240.15090.03672.7613
Viewpoint change: skip 2 viewsMulti-view DINOv2-B0.86550.58130.56800.92130.13160.04563.4025
Viewpoint change: skip 2 viewsBaseline + Geometry Feature0.87460.65860.70480.94070.15300.03642.7209

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.