Q-Flow: Query-Centric Language-Conditioned Affordance Flow Forecasting in Egocentric Scenes

Affordance Flow is an object-centric representation for task-driven future motion. However, predicting affordance flow from egocentric observations remains challenging due to occlusions, depth noise, and diverse human-object interactions. Given a single RGB-D observation, a natural-language instruction, and 3D query points sampled on the manipulated object, we propose Q-Flow, an end-to-end, query-centric predictor that forecasts multi-step 3D trajectories for these points, which serve as a geome...     «

Affordance Flow is an object-centric representation for task-driven future motion. However, predicting affordance flow from egocentric observations remains challenging due to occlusions, depth noise, and diverse human-object interactions. Given a single RGB-D observation, a natural-language instruction, and 3D query points sampled on the manipulated object, we propose Q-Flow, an end-to-end, query-centric predictor that forecasts multi-step 3D trajectories for these points, which serve as a geome...     »