PolarStream

Streaming Object Detection and Segmentation with Polar Pillars

PolarStream,
Qi Chen, Sourabh Vora, Oscar Beijbom,
NeurIPS 2021 Poster (arXiv 2006.11275)

@article{chen2021polarstream,
    title={PolarStream: Streaming Object Detection and Segmentation with Polar Pillars},
    author={Chen, Qi and Vora, Sourabh and Beijbom, Oscar},
    journal={Advances in Neural Information Processing Systems},
    volume={34},
    year={2021}
}

Contact

Any questions or suggestions are welcome!

Qi Chen qchen42@jhu.edu

Abstract

Recent works recognized lidars as an inherently streaming data source and showed that the end-to-end latency of lidar perception models can be reduced significantly by operating on wedge-shaped point cloud sectors rather then the full point cloud. However, due to use of cartesian coordinate systems these methods represent the sectors as rectangular regions, wasting memory and compute. In this work we propose using a polar coordinate system and make two key improvements on this design. First, we increase the spatial context by using multi-scale padding from neighboring sectors: preceding sector from the current scan and/or the following sector from the past scan. Second, we improve the core polar convolutional architecture by introducing feature undistortion and range stratified convolutions. Experimental results on the nuScenes dataset show significant improvements over other streaming based methods. We also achieve comparable results to existing non-streaming methods but with lower latencies.

Main results

3D detection on NuScenes test set

	MAP ↑	NDS ↑	PKL ↓	FPS ↑
PolarStream-Full Sweep	52.9	61.2		26.3
PolarStream-4 CPx1	53.5	61.8	89.3	47.2
PolarStream-4 CPx2	52.9	61.2		47.2

LiDAR Semantic Segmentation on NuScenes test set

	mIoU	freq_weighted mIoU	FPS
PolarStream-Full Sweep	73.4	87.4	33.9
PolarStream-4 CPx1	73	87.5	59.2
PolarStream-4 CPx2	73.1	87.5	59.2

Panoptic Segmentation on NuScenes test set

	PQ	SQ	RQ	FPS
PolarStream-Full Sweep	71	86	82	22.3

Panoptic Segmentation on NuScenes Validation Set (following Panoptic-PolarNet 's label generation & evaluation)

	PQ	SQ	RQ	FPS
PolarStream-Full Sweep	68.7	85.3	79.9	22.3
PolarStream-4 CPx1	69	85.2	80.4	44.3
PolarStream-4 CPx2	69.6	85.5	80.8	44.3

All results are tested on a V100 with batch size 1.

Highlighted Features

• Polar Representation
• Artificially simulating streaming lidar
• Multi-scale context padding
• Simultaneous object detection and semantic segmentation
• Single detection head with comparable accuracy to multi-group heads
• Panoptic labels and predictions generation (compatible with nuScenes official panoptic eval)
• Reimplementation of STROBE
• Reimplementation of Han et. al.
• Dynamic voxelization

Use PolarStream

Installation

Please refer to INSTALL to set up libraries needed for distributed training.

Common settings and notes

• The experiments are run with PyTorch 1.9, CUDA 11.2, and CUDNN 7.5.
• The training is conducted on 8 V100 GPUs
• Testing times are measured on a V100 GPU with batch size 1.

nuScenes 3D Detection

We provide training / validation configurations, pretrained models in the paper

Benchmark Evaluation and Training

Please refer to GETTING_START to prepare the data. Then follow the instruction there to reproduce our detection, semantic segmentation and panoptic segmentation results. Configurations are included in configs.

Model Zoo

PolarStream with PointPillars backbone

Model	det FPS	seg FPS	panoptic FPS	Test MAP	Test NDS	Test mIoU	Test freq_weigted mIoU	Validation PQ	Validation SQ	Validation RQ	Link
polarstream_det_n_seg_1_sector	26.3	33.9	22.3	52.9	61.2	73.4	87.4	68.7	85.3	79.9	URL
polarstream_det_n_seg_4_sector_bidirectional	47.2	59.2	44.3	52.9	61.2	73.1	87.5	69.6	85.5	80.8	URL

PolarStream with VoxelNet backbone (we only present full-sweep models here)

Model	val det mAP	Link
voxelnet_det_cylinder_singlehead	57.7	URL

Model	val seg mIoU	Link
voxelnet_seg_cylinder	77.7	URL

Reimplementation of STROBE and Han et. al.

Model	Link
han_1_sector	URL
han_4_sector	URL
strobe_1_sector	URL
strobe_4_sector	URL

License

PolarStream is release under MIT license (see LICENSE). It is developed based on a forked version of CenterPoint. We also incorperate code from PolarNet. See the NOTICE for details. Note that both nuScenes and Waymo datasets are under non-commercial licenses.

Acknowlegement

This project is not possible without multiple great opensourced codebases. We list some notable examples below.

• det3d
• second.pytorch
• mmcv
• mmdetection
• OpenPCDet
• polarseg
• cylinder3d