Contextual Range-View Projection for 3D LiDAR Point Clouds

Range-view projection provides an efficient method for transforming 3D LiDAR point clouds into 2D range image representations, enabling effective processing with 2D deep learning models. However, a major challenge in this projection is the many-to-one conflict, where multiple 3D points are mapped onto the same pixel in the range image, requiring a selection strategy. Existing approaches typically retain the point with the smallest depth (closest to the LiDAR), disregarding semantic relevance and object structure, which leads to the loss of important contextual information. In this paper, we extend the depth-based selection rule by incorporating contextual information from both instance centers and class labels, introducing two mechanisms: \textit{Centerness-Aware Projection (CAP)} and \textit{Class-Weighted-Aware Projection (CWAP)}. In CAP, point depths are adjusted according to their distance from the instance center, thereby prioritizing central instance points over noisy boundary and background points. In CWAP, object classes are prioritized through user-defined weights, offering flexibility in the projection strategy. Our evaluations on the SemanticKITTI dataset show that CAP preserves more instance points during projection, achieving up to a 3.1% mIoU improvement compared to the baseline. Furthermore, CWAP enhances the performance of targeted classes while having a negligible impact on the performance of other classes

💡 Research Summary

The paper tackles a fundamental limitation of range‑view projection, a popular technique for converting raw 3D LiDAR point clouds into dense 2D range images that can be processed efficiently with convolutional or transformer‑based networks. The core problem is the many‑to‑one mapping: many 3D points often fall onto the same pixel, especially when the horizontal resolution is limited or when sensor noise, ego‑motion correction, or scan‑line ordering issues are present. Conventional solutions simply keep the point with the smallest depth (closest to the sensor). While this guarantees that the visible surface is preserved, it completely ignores semantic relevance and object structure, discarding valuable contextual information such as interior points of objects or points belonging to important classes.

To address this, the authors propose two complementary, context‑aware projection strategies that are applied only during training, where ground‑truth instance masks and class labels are available. During inference the system falls back to the standard depth‑based selection, preserving real‑time performance.

1. Centerness‑Aware Projection (CAP)

   • For each annotated instance, the axis‑aligned bounding box midpoint is computed as the geometric center μ.
   • A 3‑D Gaussian function f(p) = exp(‑½‖p‑μ‖²) provides a “centerness” score for every point p belonging to that instance; points near the center receive scores close to 1, while boundary points receive lower scores.
   • The selection score for a point becomes s = depth × (1/f(p) + ε). When multiple points compete for the same pixel, the point with the smallest s is chosen. Because 1/f(p) grows for points far from the center, central points can be selected even if they are slightly farther away than a background point.
   • Points that do not belong to any instance (stuff) are assigned f(p)=0, reverting to pure depth‑based selection for those regions.

   CAP therefore biases the projection toward interior, structurally informative points of objects, reducing the loss of object shape and improving downstream semantic segmentation.

2. Class‑Weighted‑Aware Projection (CWAP)

   • Users define a weight w