Optimizing 3D Object Detection with Data Importance-Based Loss Reweighting

With the advancement of AI technology, deep learning-based intelligent driving assistance systems have seen substantial growth. However, 3D object detection remains a significant challenge due to LiDAR’s characteristics, such as sparse point clouds, varying point cloud density, and object occlusion, resulting in incomplete data. To enhance accuracy, models must be more robust. Past approaches emphasized model design, feature extraction, and obtaining finer features. In contrast, our approach introduces a novel perspective, addressing 3D object detection by focusing on sample processing without altering the model architecture. We found that point cloud variations can be substantial even within the same category. Adding such incomplete/corrupted samples to training does not improve performance; it can lead to model confusion and reduced generalization. This study proposed inferring the importance of samples based on the sample dispersed ratio and model reflection, encompassing classification and regression loss caused by sample variations. We utilize our Important Sample Selection (ISS) module to predict the sample’s importance for training and adjust the loss function to prioritize informative samples. We train and evaluate our detectors using the KITTI dataset. The experimental results show that our selection approach enhances overall detection performance without increasing parameter count.