PolarStream: Streaming Object Detection and Segmentation with Polar Pillars

Authors: Qi Chen, Sourabh Vora, Oscar Beijbom

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on the nu Scenes dataset show significant improvements over other streaming based methods. We also achieve comparable results to existing non-streaming methods but with lower latencies. Results on the nu Scenes dataset show that our proposed model Polar Stream outperforms all streaming methods in both panoptic quality and speed. Polar Stream also stays competitive with the top-performing lidar perception methods on the nu Scenes leaderboard while being at least twice as fast as the rest. We do several ablation studies and extensive analysis to show the effectiveness of Polar Stream.
Researcher Affiliation Collaboration Qi Chen Johns Hopkins University Baltimore, MD 21218 qchen42@jhu.edu Sourabh Vora Motional Santa Monica, CA 90401 sourabh.vora@motional.com Oscar Beijbom Motional Santa Monica, CA 90401 oscar.beijbom@motional.com
Pseudocode No The paper describes methods and processes (e.g., in Section 3), but it does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an explicit statement about releasing the source code for the described methodology, nor does it include a link to a code repository.
Open Datasets Yes Since there is no streaming lidar dataset available, we simulate a streaming system from the Nu Scenes dataset [3] by slicing the point clouds into n sectors according to their azimuth. [...] [3] Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuscenes: A multimodal dataset for autonomous driving. arXiv preprint arXiv:1903.11027 (2019)
Dataset Splits Yes The dataset contains 1, 000 scenes, comprising 700 scenes for training, 150 scenes for validation and 150 scenes for test.
Hardware Specification Yes All runtimes are measured on a single V100 GPU using Pytorch.
Software Dependencies No The paper mentions 'Pytorch' but does not specify its version number or any other software dependencies with their respective versions.
Experiment Setup Yes For polar pillars with n sectors per sweep, r, θ, z range is [0.3, 50.3]m, [ 3.1488, 3.1488 + 6.2976/n] rad and [ 5, 3]m, the pillar size is (0.098, 0.0123, 8). For Cartesian pillars, the pillar size is (0.2, 0.2, 8). [...] We set segmentation loss weight to 2 and classification loss to 1 for both polar pillars and Cartesian pillars. For Cartesian pillars the bounding box regression weight is 0.25. For polar pillars, since regression is harder, we set the loss weight to 0.5. [...] we conduct random flipping along x, y axes, scaling with a scale factor sampled from [0.95, 1.05], rotation around z axis between [-0.3925, 0.3925] rad and translation in range [0.2, 0.2, 0.2] m in x, y, z axis.