Leveraging Anchor-based LiDAR 3D Object Detection via Point Assisted Sample Selection

TL;DR

Experimental results demonstrate that the application of PASS elevates the average precision of anchor-based LiDAR 3D object detectors to a novel state-of-the-art, thereby proving the effectiveness of the proposed approach.

Abstract

3D object detection based on LiDAR point cloud and prior anchor boxes is a critical technology for autonomous driving environment perception and understanding. Nevertheless, an overlooked practical issue in existing methods is the ambiguity in training sample allocation based on box Intersection over Union (IoU_box). This problem impedes further enhancements in the performance of anchor-based LiDAR 3D object detectors. To tackle this challenge, this paper introduces a new training sample selection method that utilizes point cloud distribution for anchor sample quality measurement, named Point Assisted Sample Selection (PASS). This method has undergone rigorous evaluation on two widely utilized datasets. Experimental results demonstrate that the application of PASS elevates the average precision of anchor-based LiDAR 3D object detectors to a novel state-of-the-art, thereby proving the effectiveness of the proposed approach. The codes will be made available at https://github.com/XJTU-Haolin/Point_Assisted_Sample_Selection.

Figures (5)

• Figure 1: Illustration of anchor sample selection in different modalities (blue boxes indicate objects, best viewed in color). Anchor sample selection is required to assign samples into three subsets: positive, negative, and ignored. (a) Sample selection in image domain: the ambiguity of sample selection according to different measurement thresholds ($\mathcal{T}_{pos}^{}$ and $\mathcal{T}_{neg}^{}$) is small, as the sample selection metric $\mathcal{S}$ = IoU$_{box}$ primarily reflects the range of object pixel features contained by the anchor. (b) Sample selection in pointcloud domain: due to the sparsity nature of the LiDAR pointcloud, the same IoU$_{box}$-based sample selection scheme encounters greater ambiguity. For example, although anchor No.1 was selected as a positive sample with a high IoU$_{box}$, it lacked sufficient object point features compared to the ignored anchor No.2. Similarly, compared to the ignored anchor No.3, selecting anchor No.4, despite with more object point features, as a negative sample solely due to its lower IoU${box}$ is ambiguous. (c) To mitigate the ambiguity of sample selection in pointcloud domain, the goal is to design a novel form of the sample selection metric ($\mathcal{S^{\prime}}$) that better suits the learning of object point features.
• Figure 2: Statistics of training anchor samples distribution over the KITTI dataset. (a) The scatter diagram of the relation between IoU$_{box}$ and IoU$_{point}$. (b) The histogram diagram of IoU$_{point}$ of those ignored anchor samples determined by IoU$_{box}$-based sample selection scheme. (c) The histogram of IoU$_{point}$ of those negative anchor samples determined by IoU$_{box}$-based sample selection scheme. (d) The histogram of IoU$_{point}$ of those positive anchor samples determined by IoU$_{box}$-based sample selection scheme.
• Figure 3: Visualized case studies of ambiguity associated with IoU$_{box}$-based sample selection (best viewed in color). The green boxes and blue boxes represent the anchor samples and ground truths respectively. Points in the intersection of the anchor sample and ground truth are indicated as red points. Points that belong to ground truth but do not belong to the intersection are indicated as blue points. Points that belong to the anchor but do not belong to the intersection are represented as green points.
• Figure 4: Qualitative results on training sample selection (best viewed in color). $\mathcal{S}$ and $\mathcal{S}^{\prime}$ are calculated by the IoU$_{box}$-based and the PASS-based sample selection measurement, respectively. The green boxes and blue boxes represent the anchor samples and ground truths respectively. Points in the intersection of the anchor sample and ground truth are indicated as red points. Points that belong to ground truth but do not belong to the intersection are indicated as blue points. Points that belong to the anchor but do not belong to the intersection are represented as green points, and the rest of the background points are represented as black points.
• Figure 5: Qualitative results on challenging scenes from the KITTI test set (best viewed in color). The top row displays the reference images, and the bottom row shows the corresponding LiDAR pointcloud with detection results from the proposed PASS-PV-RCNN++. The green, blue, and yellow boxes represent the prediction of car, pedestrian, and cyclist respectively. Certain cases that are worth further discussing (see Section \ref{['qualitative-analysis']}) are highlighted by zoomed-in.