You Yang ; Shi Ruoxi ; Wang Weiming ; Lu Cewu

Abstract

In this paper, we tackle the problem of category-level 9D pose estimation inthe wild, given a single RGB-D frame. Using supervised data of real-world 9Dposes is tedious and erroneous, and also fails to generalize to unseenscenarios. Besides, category-level pose estimation requires a method to be ableto generalize to unseen objects at test time, which is also challenging.Drawing inspirations from traditional point pair features (PPFs), in thispaper, we design a novel Category-level PPF (CPPF) voting method to achieveaccurate, robust and generalizable 9D pose estimation in the wild. To obtainrobust pose estimation, we sample numerous point pairs on an object, and foreach pair our model predicts necessary SE(3)-invariant voting statistics onobject centers, orientations and scales. A novel coarse-to-fine votingalgorithm is proposed to eliminate noisy point pair samples and generate finalpredictions from the population. To get rid of false positives in theorientation voting process, an auxiliary binary disambiguating classificationtask is introduced for each sampled point pair. In order to detect objects inthe wild, we carefully design our sim-to-real pipeline by training on syntheticpoint clouds only, unless objects have ambiguous poses in geometry. Under thiscircumstance, color information is leveraged to disambiguate these poses.Results on standard benchmarks show that our method is on par with currentstate of the arts with real-world training data. Extensive experiments furthershow that our method is robust to noise and gives promising results underextremely challenging scenarios. Our code is available onhttps://github.com/qq456cvb/CPPF.