False Positive Sampling-based Data Augmentation for Enhanced 3D Object Detection Accuracy

TL;DR

The paper addresses false positives arising from ground-truth sampling in 3D LiDAR object detection. It introduces false-positive (FP) sampling, which builds a FP sample database from the model’s predictions and integrates FP and GT samples during training to sharpen the decision boundary. Across KITTI and Waymo, and over multiple architectures, FP sampling reduces false positives and improves overall detection accuracy, with notable gains in pedestrian detection. The approach is model- and dataset-agnostic but incurs computational overhead and additional hyperparameters, suggesting directions for efficiency improvements and automated tuning.

Abstract

Recent studies have focused on enhancing the performance of 3D object detection models. Among various approaches, ground-truth sampling has been proposed as an augmentation technique to address the challenges posed by limited ground-truth data. However, an inherent issue with ground-truth sampling is its tendency to increase false positives. Therefore, this study aims to overcome the limitations of ground-truth sampling and improve the performance of 3D object detection models by developing a new augmentation technique called false-positive sampling. False-positive sampling involves retraining the model using point clouds that are identified as false positives in the model's predictions. We propose an algorithm that utilizes both ground-truth and false-positive sampling and an algorithm for building the false-positive sample database. Additionally, we analyze the principles behind the performance enhancement due to false-positive sampling. Our experiments demonstrate that models utilizing false-positive sampling show a reduction in false positives and exhibit improved object detection performance. On the KITTI and Waymo Open datasets, models with false-positive sampling surpass the baseline models by a large margin.

Figures (6)

• Figure 1: Comparison of data augmentation methods in 3D object detection: (a) the original point cloud without any data augmentation via GT or FP sampling, (b) data augmentation via GT sampling, and (c) data augmentation via GT and FP sampling. Blue boxes represent ground truths, while green boxes indicate false positives.
• Figure 2: The distribution of false positives with or without GT sampling based on the model's confidence score. The model trained with GT sampling tends to be over-confident, if there's not a appropriate constraints to model's prediction.
• Figure 3: The proposed data augmentation and model training process, including the FP sample database management
• Figure 4: The decision boundaries of the model changed by different data augmentation methods. Blue circles represent positive samples, while red triangles represent negative samples. Applying GT and FP sampling for data augmentation can modify the model's decision boundary (black dotted line) to be closer to the actual decision boundary (yellow solid line). Note that the figures result from dimensionality reduction from a very high dimensional space into two dimensions.
• Figure 5: Comparison of GT sampling and GT & FP sampling results on the KITTI validation dataset for the SECOND model. Pink and green bounding boxes represent the predictions made by the models trained with GT sampling and GT & FP sampling methods, respectively.