Research Project · 3D Perception · Zero-Shot

Geometry-Aware
Zero-Shot 3D Affordance Mapping

Combining CLIP semantic features with query-conditioned geometric priors to localize interaction regions in 3D point clouds — no training data required. Evaluated on the LASO benchmark (CVPR 2025).

CLIP-only methods lift 2D patch features into 3D and score points by text similarity. We multiply this semantic score by a geometry score derived from local surface descriptors — encoding the physical common sense that, e.g., graspable surfaces have consistent normals and slight curvature, while placeable surfaces are flat and upward-facing.

final_score(p) = √( semantic_score(p) × geometric_score(p) )
CLIP lifting
2D patches → 3D
×
Geometric prior
PCA normals + query filter
Affordance map
Per-point score [0,1]
Geometry improves region quality, not just peak detection.

CLIP-only achieves higher P@50 (finds the right single points), but our geometry-aware fusion wins on IoU and SIM across all four affordance categories — producing better-shaped regions. For robotics this matters: a robot needs a graspable surface, not just the highest-scoring point.

Multiplicative fusion outperforms additive alternatives.

Multiplying semantic and geometric scores enforces that both must agree — a point scores high only if it both looks right (CLIP) and is shaped right (geometry). Additive fusion lets high semantic scores compensate for geometrically implausible regions, which degrades IoU.

Short queries outperform descriptive ones on P@50 (novel finding).

For the support affordance, "region for support" achieves P@50=0.960 vs. 0.340 for the full descriptive query — while IoU is stable across all phrasings. This suggests CLIP's text encoder is sensitive to query verbosity for peak localization, but geometry provides stable signal regardless. Practical implication: use short action-centric queries.

CLIP only baseline
Grasp IoU 0.262
Support IoU0.205
Contain IoU0.375
Move IoU 0.332
Grasp P@50 0.382
Geometry only
Grasp IoU 0.295
Support IoU0.127
Contain IoU0.347
Move IoU 0.452
Grasp P@50 0.332
Geometry-aware ours
Grasp IoU 0.354
Support IoU0.258
Contain IoU0.541
Move IoU 0.509
Grasp P@50 0.323
IoU by affordance
SIM by affordance
Variant P@50 IoU SIM
(A) CLIP only 0.382 0.262 0.261
(B) Geometry only 0.332 0.295 0.292
(C) Additive fusion 0.327 0.294 0.291
(D) Multiplicative (ours) 0.323 0.354 0.325
(E) Harmonic fusion 0.329 0.353 0.328
Query P@50 IoU

Short action-centric queries dramatically outperform verbose descriptions on P@50, while IoU remains stable — suggesting geometry provides robustness to prompt phrasing.

Hand-specified priors
Geometric priors are manually defined per affordance category. Learning them from a small ShapeNet set would remove hand-engineering while preserving zero-shot generalization at test time.
CLIP feature resolution
ViT-L/14 produces ~14px patches. Affordance regions smaller than ~14px in projection are underresolved. DINOv2 with 14px stride gives better spatial resolution for small objects.
P@50 vs IoU trade-off
Our method trades peak precision for better region quality. For single-point grasping tasks CLIP-only may suffice; for surface coverage (bi-manual grasping) geometry-aware fusion is preferable.
Non-metric depth
On webcam input, depth is estimated by Depth Anything V2 — geometrically consistent but not metric. Normal estimation quality degrades slightly without true depth.