GANOCS: Domain Adaptation of Normalized Object Coordinate Prediction Using Generative Adversarial Training
Bahaa Aldeeb, Sahith Reddy Chada, and Karthik Desingh
In First Workshop on Out-of-Distribution Generalization in Robotics at CoRL, 2023
Estimating 2D-3D correspondences has proven to be a very useful tool for category-level pose and scale estimation and robot manipulation tasks; how- ever, it is hindered by the difficulty of obtaining 3D object models and labels. Simulation reduces the burden of labeling but introduces a gap between the train- ing and operational domains. We introduce a novel architecture for integrating cross-domain data in the training of NOCS predictors (a form of 2D-3D corre- spondences). We leverage Generative-Adversarial-Networks (GANs) to avoid the need for burdensome real-data labeling by using a domain-agnostic discrimina- tor as a supervisor. This work presents results demonstrating the potential of our method.
