2022/03/17 by Alexey Bokhovkin, Bokhovkin, Alexey, Angela Dai +1 · 1 citation
Engineering · Computer Science · #3D Shape Modeling and Analysis #Advanced Neural Network Applications #Human Pose and Action Recognition
paper · pdf · doi:10.48550/arxiv.2203.09375
3D object recognition has seen significant advances in recent years, showing\nimpressive performance on real-world 3D scan benchmarks, but lacking in object\npart reasoning, which is fundamental to higher-level scene understanding such\nas inter-object similarities or object functionality. Thus, we propose to\nleverage large-scale synthetic datasets of 3D shapes annotated with part\ninformation to learn Neural Part Priors (NPPs), optimizable spaces\ncharacterizing geometric part priors. Crucially, we can optimize over the\nlearned part priors in order to fit to real-world scanned 3D scenes at test\ntime, enabling robust part decomposition of the real objects in these scenes\nthat also estimates the complete geometry of the object while fitting\naccurately to the observed real geometry. Moreover, this enables global\noptimization over geometrically similar detected objects in a scene, which\noften share strong geometric commonalities, enabling scene-consistent part\ndecompositions. Experiments on the ScanNet dataset demonstrate that NPPs\nsignificantly outperforms state of the art in part decomposition and object\ncompletion in real-world scenes.\n