Source-free Unsupervised Domain Adaptation for 3D Object Detection in Adverse Weather

A domain shift exists between the distributions of large scale, outdoor lidar datasets due to being captured using different types of lidar sensors, in different locations, and under varying weather conditions. Inclement weather in particular affects the quality of lidar data, adding artifacts such as scattered and missed points, leading to a drop in performance of 3D object detection networks trained on standard lidar datasets. Domain adaptation methods seek to adapt source-trained neural networks to a target domain. Pseudo-label based self training approaches are popular methods for source-free unsupervised domain adaptation. However, their efficacy depends on the quality of the labels generated by the source trained model. These labels may be incorrect with high confidence, rendering thresholding methods ineffective. In order to avoid reinforcing errors caused by label noise, we propose an uncertainty-aware mean teacher framework which implicitly filters incorrect pseudo-labels during training. Leveraging model uncertainty allows the mean teacher network to perform implicit filtering by down-weighing losses corresponding to uncertain pseudo-labels. Effectively, we perform automatic soft-sampling of pseudo-labeled data while aligning predictions from the student and teacher networks. We demonstrate our domain adaptation method on an adverse weather dataset created by augmenting lidar scenes from KITTI with rain, snow, and fog and show that it out-performs current domain adaptation frameworks. We make our code publicly available ¹¹https://github.com/deeptibhegde/UncertaintyAwareMeanTeacher.

I. Introduction

Perception is an important part of autonomous driving systems, with navigation and decision making relying heavily on the ability of the vehicle to correctly localize and classify the objects around it. Recent pure lidar-based 3D object detectors have proven to perform extremely well on large public datasets and have topped their challenge leaderboards [1]–[3]. Although containing similar scenes of roads, pedestrians, and vehicles, they tend to differ from each other in terms of pointcloud density, the average size of lanes, as well as the types of vehicles present [4]. This is due to the fact that these datasets are collected using different types of lidar sensors in varying locations around the world, and at times, under varying weather conditions. In weather scenarios such as rain, snow and fog, data from lidar sensors may get corrupted due to reduced signal-to-noise ratio (SNR) or scattered power from the droplets and particles in the air.