Registration between Point Cloud Streams and Sequential Bounding Boxes via Gradient Descent

TL;DR

The paper addresses the challenge of registering sequential 3D bounding boxes to evolving point-cloud streams without heavy reliance on annotated data. It introduces a differentiable objective that combines closeness, enclosure, smoothness, and alignment losses ($L_c$, $L_e$, $L_s$, $L_a$) and optimizes the bounding-box parameters directly using Newton-type methods with LM-BFGS acceleration. This optimization-based, ICP-inspired approach aims to provide an explainable and configurable alternative to learning-based methods, achieving up to 40% IoU gains in simulated 2D and 3D BEV tasks. The work demonstrates that carefully designed objective terms and optimization strategies can yield robust registration in the presence of occlusion and trajectory constraints, with potential applicability to autonomous driving and 3D auto-labeling, while also outlining limitations for high-dimensional parameter spaces and real-world data.

Abstract

In this paper, we propose an algorithm for registering sequential bounding boxes with point cloud streams. Unlike popular point cloud registration techniques, the alignment of the point cloud and the bounding box can rely on the properties of the bounding box, such as size, shape, and temporal information, which provides substantial support and performance gains. Motivated by this, we propose a new approach to tackle this problem. Specifically, we model the registration process through an overall objective function that includes the final goal and all constraints. We then optimize the function using gradient descent. Our experiments show that the proposed method performs remarkably well with a 40\% improvement in IoU and demonstrates more robust registration between point cloud streams and sequential bounding boxes

Figures (5)

• Figure 1: Point cloud streams (point clouds of "car" at different times) and their corresponding sequential bounding boxes (red boxes). The registration task is to align sequential bounding boxes to the given point cloud streams.
• Figure 2: The sketch diagram about the landscape of Enclosure$\mathbbmsl{L}_e$ with $\textit{L1}$ and $\textit{L2}$-norm.
• Figure 3: The comparison registration results. In Fig. (a), the ground truth bounding boxes are depicted in Red color, the initial bounding box is green, and the registration bounding box is blue color. Fig. (b) displays the error value. All point clouds and bounding box are temporarily down-sampled by 5 to make plotting clean.
• Figure 4: The comparison registration results. In Fig. (a), the ground truth bounding boxes are depicted in Red color, the initial bounding box is green, and the registration bounding box is the blue color. Fig. (b) displays the error value.
• Figure 5: A visualized point cloud example of an SUV for the experiment. We give the (a) left-right and (b) top-down views of the example.