Unsupervised Domain Adaptation for 3D Object Detection Using Adversarial Adaptation : Learning Transferable LiDAR Features for a Delivery Robot

Sammanfattning: 3D object detection is the task of detecting the full 3D pose of objects relative to an autonomous platform. It is an important perception system that can be used to plan actions according to the behavior of other dynamic objects in an environment. Due to the poor generalization of object detectors trained and tested on different datasets, this thesis concerns the utilization of unsupervised domain adaptation to train object detectors fit for mobile robotics without any labeled training data. To tackle the problem a novel approach Unsupervised Adversarial Domain Adaptation 3D (UADA3D) is presented to adapt LiDAR-based detectors, through drawing inspiration from the success of adversarial adaptation for 2D object detection in RGB images. The method adds learnable discriminator layers that discriminate between the features and bounding box predictions in the labeled source and unlabeled target data. The gradients are then reversed through gradient reversal layers during backpropagation to the base detector, which in turn learns to extract features that are similar between the domains in order to fool the discriminator. The method works for multi-class detection by simultaneous adaptation of all classes in an end-to-end trainable network and works for both point-based and voxel-based single-stage detectors. The results show that the proposed method increases detection scores for adaptation from dense to sparse point clouds and from simulated data toward the data of a mobile delivery robot, successfully handling the two relevant domain gaps given by differences in marginal and conditional probability distributions.