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

Key points

• One challenge is the query-target gap: a semantic discrepancy between query features encoding interaction intents and target-region features, leading to inaccurate grounding.1
• To address this issue, we propose BridgeAfford, a framework that reduces the query-target semantic gap from two complementary perspectives.1
• First, Interaction Region Alignment (IRA) promotes cross-object alignment among geometrically similar interaction regions during training.1
• Second, Visual Evidence Aggregation (VEA) leverages a 2D affordance model to refine the attention distribution and guide the query toward the correct region, yielding more target-aware queries.1

Sources (1)

• [1]BridgeAfford: Bridging Query-Target Gap for Generalizable 3D Affordance Grounding
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 02:03 PM

  Language-guided 3D affordance grounding aims to localize functional regions on 3D objects according to interaction intents.

  One challenge is the query-target gap: a semantic discrepancy between query features encoding interaction intents and target-region features, leading to inaccurate grounding.

Extractive summary: sentences quoted from the sources.