3D Object Detection from PointClouds is an essential part of the perception stack of autonomous vehicles. Although LiDAR is a commonly used sensor for detection, it has a larger domain shift from dataset to dataset, compared to other sensor modalities. For this reason, collecting large amounts of ground truth data is crucial for adapting an off-the-shelf model to the target domain. Since the cost of labeling is an important consideration for companies, self-supervision is a good approach to overcome the problem. In this research, we introduce a self-supervision and auto-labeling pipeline that can be executed to adopt to a target domain, without any amount of training in the target domain at all. The code is publicly available at https://github.com/yucedagonurcan/TRADE_TrackingToDetect
- We show the introduction of the detections with prior tracker information helps the accuracy of the psuedo-labeled dataset
- We introduce a dynamic clustering algorithm that can optimize the found object with respect the tracklet and its historic data
- We show adaptively selecting the proposal space of the shape fitting by the given tracklet information both reduces computation complexity and improves the output object accuracy
- We present simple yet effective object fusion method that both keeps the best proposal for a single object while effectively introduces the lost objects and increasing the recall in overall pipeline
- We carefully lay out the importance of the life-cycle management for MOT modules and show that with the correct scoring and life-cycle management of the tracklets, we further improve the overall accuracy of the pipeline
