Attentive Prototypes for Source-free Unsupervised Domain Adaptive 3D Object Detection

3D object detection networks tend to be biased towards the data they are\ntrained on. Evaluation on datasets captured in different locations, conditions\nor sensors than that of the training (source) data results in a drop in model\nperformance due to the gap in distribution with the test (or target) data.\nCurrent methods for domain adaptation either assume access to source data\nduring training, which may not be available due to privacy or memory concerns,\nor require a sequence of lidar frames as an input. We propose a single-frame\napproach for source-free, unsupervised domain adaptation of lidar-based 3D\nobject detectors that uses class prototypes to mitigate the effect pseudo-label\nnoise. Addressing the limitations of traditional feature aggregation methods\nfor prototype computation in the presence of noisy labels, we utilize a\ntransformer module to identify outlier ROI's that correspond to incorrect,\nover-confident annotations, and compute an attentive class prototype. Under an\niterative training strategy, the losses associated with noisy pseudo labels are\ndown-weighed and thus refined in the process of self-training. To validate the\neffectiveness of our proposed approach, we examine the domain shift associated\nwith networks trained on large, label-rich datasets (such as the Waymo Open\nDataset and nuScenes) and evaluate on smaller, label-poor datasets (such as\nKITTI) and vice-versa. We demonstrate our approach on two recent object\ndetectors and achieve results that out-perform the other domain adaptation\nworks.\n