Core Concepts
Controllable generative model for 3D articulated objects enhances realism and user control.
Abstract
Introduction:
Articulated objects are common in real-world scenarios.
Challenges in generating 3D articulated objects controllably.
Method:
Utilizes denoising diffusion-based method with attention modules.
Takes object category label and part connectivity graph to generate geometry and motion parameters.
Experiments:
Outperforms state-of-the-art in generating realistic articulated objects.
Better compatibility with user-specified constraints.
Ablations:
Removal of attention modules leads to lower quality objects.
Failure Cases:
Overlapping parts, mismatched joint axes, unrealistic retrieved parts identified as limitations.
Stats
"Our experiments show that our method outperforms the state-of-the-art in articulated object generation."
"Our method also demonstrates better compatibility with various conditional input scenarios enabling better user-controlled generation."
Quotes
"Noisy sample xt is obtained by gradually adding Gaussian noise ϵ ∼ N(0, I)."
"Our experiments show that this abstraction coupled with a series of appropriately designed attention modules improve joint modeling of parts and motion compared to prior work."