@inproceedings{raptors2026-c049,
  title     = {{LiDAR Challenges in Camera-Centric Perception}},
  author    = {Nagvekar, Anushree},
  booktitle = {Proceedings of the Raptors Conference},
  volume    = {1},
  year      = {2026},
  publisher = {Hackathon Raptors},
  address   = {London, United Kingdom},
  doi       = {10.68495/rc.2026.C049},
  url       = {https://doi.org/10.68495/rc.2026.C049},
  abstract  = {LiDAR contributes valuable three-dimensional geometric information to perception, and in camera-centric systems it also brings difficulties: sparse point clouds, limited semantic detail, sensitivity to weather, and the complexity of fusing it with other modalities. This paper examines that tension directly. The same sparsity that makes semantic segmentation harder leaves the structured geometry useful elsewhere, improving tracking through better depth estimation, object separation and temporal consistency, so the sensor is neither straightforwardly helpful nor straightforwardly costly. The consequence for system design is that effective fusion tends to rely on selective candidate processing and adaptive attention rather than dense alignment across all modalities, since aligning everything spends compute on regions where LiDAR adds nothing. Readers take away an account of where LiDAR genuinely strengthens a camera-centric perception stack, and how to use it efficiently enough to lower the compute and bandwidth a perception system requires.}
}