BridgeAfford: Bridging Query-Target Gap for Generalizable 3D Affordance Grounding

1 INTRODUCTION

• On unseen objects, interaction-region representations may shift from training examples, making learned intent-to-region associations unreliable and widening the query–target semantic gap.Existing approaches extract target-region information into queries before predicting affordance masks from query-to-point-feature similarity.Source: §1 · p. 1
• BridgeAfford achieves leading performance on LASO and PIAD, particularly on unseen categories and objects.The introduction reports that extensive experiments validate its generalization ability.Source: §1 · pp. 2–3
• Interaction Region Alignment (IRA) encourages consistent representations for geometrically similar interaction regions across objects, providing transferable cues for target recognition.BridgeAfford uses geometrically diverse anchors and affinity-based contrastive learning to align regions to anchors by geometric similarity.Source: §1 · p. 2
• Visual Evidence Aggregation (VEA) lifts multi-view 2D affordance predictions into dense spatial evidence that guides target-region feature extraction and produces target-aware queries.The introduction frames 2D guidance as a complement for novel interaction structures that cannot be reliably matched by geometric similarity alone.Source: §1 · pp. 2–3

3D Affordance Learning

• Recent 3D affordance methods use external models, cross-modal designs, or enhanced training strategies, yet generalization to unseen objects remains challenging.The area progressed from point-wise multi-label segmentation to interaction images or language; LASO supports sentence-level instructions.Source: §2 · p. 3

Object Representation Learning for 3D Affordance

• Unlike methods that rely on explicit affordance labels, BridgeAfford aligns local regions with similar geometric patterns, avoiding forced alignment of geometrically diverse regions for the same intent.The related-work discussion notes that extracting explicit labels with LLMs is costly and does not resolve this diversity issue.Source: §2 · p. 3

Query Enhancement for 3D Affordance

• BridgeAfford uses dense, smooth point-wise guidance from rendered-view affordance masks, unlike sparse keypoints that capture partial target-mask evidence.The discussion also contrasts this spatial guidance with abstract LLM- or MLLM-inferred semantics, which remain distant from target-region features and add computational overhead.Source: §2 · p. 3

3.1 THE BACKBONE OF BRIDGEAFFORD

• The backbone fuses point-cloud geometry with text, forms target-aware queries by aggregating target-region features, and predicts point-wise affordance by query-feature similarity.Each output value represents a point’s interaction likelihood.Source: §3.1 · pp. 3–4

3.2 INTERACTION REGION ALIGNMENT

• CAIM stores paired geometric and fused region features; DAI preserves K diverse entries by merging the most similar pair when full, while ABCL aligns each sample to its geometry-matched entry.ABCL uses geometrically similar regions as positives rather than forcing alignment based only on shared interaction function.Source: §3.2 · pp. 4–5

3.3 VISUAL EVIDENCE AGGREGATION

• VEA converts multi-view 2D affordance predictions into dense 3D visual evidence; CAMVA confidence-weights pointwise scores based on depth consistency and visibility, and VEA injects the evidence as an additive query-to-point attention bias.Source: §3.3 · pp. 5–6

3.4 TRAINING STRATEGY AND OBJECTIVE

• Training first stabilizes object representations with segmentation loss, then initializes CAIM, and finally jointly optimizes segmentation and alignment losses while updating memory online; inference omits CAIM.The 2D affordance model is trained with view-level supervision and frozen before producing heatmaps for VEA.Source: §3.4 · p. 6

4.1 EXPERIMENTAL SETUP

• On LASO, both seen and unseen settings are evaluated; unseen object-affordance pairs are held out during training, and the reported metrics are aIoU, AUC, SIM, and MAE.The implementation uses Uni3D and RoBERTa encoders, six rendered views, and a CAIM with 16 entries.Source: §4.1 · pp. 6–7

4.2 QUANTITATIVE RESULTS

• On LASO, BridgeAfford leads across all metrics on unseen objects while remaining competitive on seen objects.The evaluation reports aIoU, AUC, SIM, and MAE.Source: §4.2 · p. 7, Table 1 · p. 7, §4.1 · p. 7
• On PIAD, BridgeAfford is best among language-guided methods in both settings, improves aIoU over ViSPLA from 9.20 to 14.50, and leads reported cross-dataset metrics.For cross-dataset evaluation, models trained on LASO are evaluated directly on the PIAD test set.Source: §4.2 · pp. 7–8, Table 2 · p. 7, Table 3 · p. 8
• With 60% of LASO unseen training data, BridgeAfford surpasses all full-data baselines; with 20%, it achieves the best aIoU.Figure 3 compares BridgeAfford at different training-data ratios with baselines trained on the full set.Source: §4.2 · p. 8, Figure 3 · p. 8

4.3 ABLATION STUDIES

• Both IRA and VEA improve the backbone, especially on unseen objects, and their combination achieves the best overall performance.IRA aligns geometrically similar regions across objects; VEA supplies dense target-region evidence to guide queries.Source: §4.3 · p. 8, Table 4 · p. 8
• Geometry-based alignment outperforms label-based alignment, while Diversity-Aware Initialization outperforms random initialization, which can cause training collapse.Regions sharing an affordance label may have different geometries, and early alignment to uninformative memory entries can corrupt representation learning.Source: §4.3 · p. 9, Table 5 · p. 9
• Removing normal-based coloring or depth confidence degrades performance, while 2D-only masks are insufficient because viewpoints and occlusions disrupt smooth 3D masks.Depth confidence helps account for occlusion during multi-view aggregation; normal-based coloring distinguishes object surfaces and parts.Source: §4.3 · p. 9, Table 5 · p. 9

4.4 QUALITATIVE RESULTS

• On the LASO unseen split, BridgeAfford localizes interaction regions more accurately, with higher-confidence predictions and cleaner, more complete masks than prior methods.The qualitative comparison also shows noisy responses suppressed in irrelevant regions.Source: §4.4 · p. 9, Figure 4 · p. 10

• BridgeAfford narrows the query–target semantic gap to improve 3D affordance grounding on unseen objects.The conclusion attributes cross-object generalization to transferable local geometric structures and valuable query-localization guidance from 2D visual evidence.Source: §5 · p. 10
• Experiments on LASO and PIAD support BridgeAfford’s superiority and the potential of combining geometric structures with 2D visual priors.Source: §5 · p. 10

• The analysis varies IRA and VEA hyperparameters, including memory size, warm-up epochs, loss and attention-bias weights, EMA coefficient, and rendered views.Source: A HYPERPARAMETER ANALYSIS · p. 13
• Hyperparameters are selected on the LASO unseen validation split, while the main results are evaluated on the corresponding test split.Source: A HYPERPARAMETER ANALYSIS · p. 13

• Query–target similarity is the average cosine similarity between each query and a ground-truth-affordance-weighted pooled target-region representation.The analysis also measures similarity to non-target-region features.Source: B QUANTITATIVE ANALYSIS OF THE… · p. 13, Table 8 · p. 14
• Both IRA and VEA improve query–target similarity, and their combination achieves the highest target-region similarity.Table 8 reports query–region cosine similarity and defines Δ as target similarity minus background similarity.Source: B QUANTITATIVE ANALYSIS OF THE… · p. 14, Table 8 · p. 14

C VISUAL ANALYSIS OF DIVERSITY-AWARE INITIALIZATION

• The 16 initialized CAIM geometric features are widely distributed in t-SNE space, indicating that DAI captures diverse geometric patterns.The visualization compares the 16 anchor features (red stars) with target-region geometric features (gray points) from category-balanced samples.Source: C VISUAL ANALYSIS OF DIVERSITY-AWARE… · p. 14, Figure 5 · p. 14

D.1 ADVANTAGES OF NORMAL-BASED GAUSSIAN RENDERER

• NBGR renderings distinguish surface regions and preserve component boundaries, providing richer spatial cues for the 2D affordance model.Figure 6 visualizes renderings generated by the Normal-Based Gaussian Renderer.Source: §D.1 · p. 14, Figure 6 · p. 15

D.2 GENERALIZATION CAPABILITY OF THE 2D AFFORDANCE MODEL VERSUS THE 3D BACKBONE

• The 2D affordance model achieves better performance than the 3D backbone on the LASO unseen split, supporting its use as a visual prior.Table 9 compares their generalization capabilities on that split.Source: §D.2 · p. 15, Table 9 · p. 15

D.3 DENSE POINT-WISE EVIDENCE VS. SPARSE KEYPOINT SAMPLING

• CAMVA consistently outperforms sparse keypoint sampling across all metrics, with dense point-wise confidence providing more effective spatial guidance.Both strategies use the same backbone, 2D model, multi-view predictions, and training settings; they differ in evidence aggregation and incorporation.Source: §D.3 · p. 16, Table 10 · p. 16

D.4 EFFECT OF VISUAL EVIDENCE AGGREGATION ON QUERY ATTENTION

• With VEA, query attention becomes more concentrated on target regions and suppresses irrelevant background responses, unlike unguided attention on unseen objects.Figure 7 compares attention maps with and without visual evidence guidance.Source: §D.4 · p. 16, Figure 7 · p. 16

E FINE-GRAINED RESULTS BY OBJECT–AFFORDANCE COMBINATION

• Wrap-grasping remains challenging, particularly for mugs, bowls, and bottles, amid annotation ambiguity and extended contact surfaces.The analysis also notes that limited visibility across rendered views may leave parts of these surfaces without sufficient visual evidence.Source: E FINE-GRAINED RESULTS BY… · p. 17
• Bowl-wrap-grasp and bottle-wrap-grasp pair high SIM scores (0.745 and 0.703) with low aIoU scores (4.50 and 7.09), indicating distributional agreement without precise spatial overlap.The per-class results cover LASO unseen split object–affordance combinations and report aIoU, AUC, SIM, and MAE.Source: E FINE-GRAINED RESULTS BY… · p. 17, Table 11 · p. 17
• The analysis identifies room to improve spatial coverage and handle ambiguity in interaction regions.Source: E FINE-GRAINED RESULTS BY… · p. 17