Online Segmentation of LiDAR Sequences: Dataset and Algorithm
This work addresses real-time segmentation challenges for autonomous vehicles, offering a dataset and algorithm that significantly reduce latency and model size, though it is incremental in improving existing methods.
The authors tackled the problem of high latency in LiDAR sequence segmentation for autonomous vehicles by introducing HelixNet, a 10 billion point dataset with fine-grained labels, and Helix4D, a spatio-temporal transformer architecture that reduces latency by over 5x and model size by 50x while maintaining accuracy.
Roof-mounted spinning LiDAR sensors are widely used by autonomous vehicles. However, most semantic datasets and algorithms used for LiDAR sequence segmentation operate on $360^\circ$ frames, causing an acquisition latency incompatible with real-time applications. To address this issue, we first introduce HelixNet, a $10$ billion point dataset with fine-grained labels, timestamps, and sensor rotation information necessary to accurately assess the real-time readiness of segmentation algorithms. Second, we propose Helix4D, a compact and efficient spatio-temporal transformer architecture specifically designed for rotating LiDAR sequences. Helix4D operates on acquisition slices corresponding to a fraction of a full sensor rotation, significantly reducing the total latency. Helix4D reaches accuracy on par with the best segmentation algorithms on HelixNet and SemanticKITTI with a reduction of over $5\times$ in terms of latency and $50\times$ in model size. The code and data are available at: https://romainloiseau.fr/helixnet