Semi-supervised Implicit Scene Completion from Sparse LiDAR

New Repo

Paper

Created by Pengfei Li, Yongliang Shi, Tianyu Liu, Hao Zhao, Guyue Zhou and YA-QIN ZHANG from Institute for AI Industry Research(AIR), Tsinghua University.

For complete video, click HERE.

We use the proposed model trained on the KITTI dataset to predict implicit completion results on the novel DAIR-V2X dataset. The results are impressive:

Introduction

Recent advances show that semi-supervised implicit representation learning can be achieved through physical constraints like Eikonal equations. However, this scheme has not yet been successfully used for LiDAR point cloud data, due to its spatially varying sparsity.

In this repository, we develop a novel formulation that conditions the semi-supervised implicit function on localized shape embeddings. It exploits the strong representation learning power of sparse convolutional networks to generate shape-aware dense feature volumes, while still allows semi-supervised signed distance function learning without knowing its exact values at free space. With extensive quantitative and qualitative results, we demonstrate intrinsic properties of this new learning system and its usefulness in real-world road scenes. Notably, we improve IoU from 26.3% to 51.0% on SemanticKITTI. Moreover, we explore two paradigms to integrate semantic label predictions, achieving implicit semantic completion. Codes and data are publicly available.

Citation

If you find our work useful in your research, please consider citing:

@misc{li2021semisupervised,
    title={Semi-supervised Implicit Scene Completion from Sparse LiDAR}, 
    author={Pengfei Li and Yongliang Shi and Tianyu Liu and Hao Zhao and Guyue Zhou and Ya-Qin Zhang},
    year={2021},
    eprint={2111.14798},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Installation

Requirements

CUDA=11.1
python>=3.8
Pytorch>=1.8
numpy
ninja
MinkowskiEngine
tensorboard
pyyaml
configargparse
scripy
open3d
h5py
plyfile
scikit-image

Clone the repository:

git clone https://github.com/OPEN-AIR-SUN/SISC.git

Data preparation

Download the SemanticKITTI dataset from HERE. Unzip it into the same directory as SISC.

Training and inference

The configuration for training/inference is stored in opt.yaml, which can be modified as needed.

Scene Completion

Run the following command for a certain task (train/valid/visualize):

CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --nproc_per_node=1 main_sc.py --task=[task] --experiment_name=[experiment_name]

Semantic Scene Completion

SSC option A

Run the following command for a certain task (ssc_pretrain/ssc_valid/train/valid/visualize):

CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --nproc_per_node=1 main_ssc_a.py --task=[task] --experiment_name=[experiment_name]

Here, use ssc_pretrain/ssc_valid to train/validate the SSC part. Then the pre-trained model can be used to further train the whole model.

SSC option B

Run the following command for a certain task (train/valid/visualize):

CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --nproc_per_node=1 main_ssc_b.py --task=[task] --experiment_name=[experiment_name]

Model Zoo

Our pre-trained models can be downloaded here:

Ablation	Pretrained Checkpoints
data augmentation	no aug	rotate & flip
Dnet input	radial distance	radial distance & height
Dnet structure	last1 pruning	last2 pruning	last3 pruning	last4 pruning	Dnet relu	4convs output
Gnet structure	width128 depth4	width512 depth4	width256 depth3	width256 depth5	Gnet relu
point sample	on:off=1:2	on:off=2:3
positional encoding	no encoding	incF level10	incT level5	incT level15
sample strategy	nearest
scale size	scale 2	scale 4	scale 8	scale 16	scale 32
shape size	shape 128	shape 512
SSC	SSC option A	SSC option B
Seg-Net&SSC-Net for SSC-A	ssc_pretrain

These models correspond to the ablation study in our paper. The Scale 4 works as our baseline.

Ablation	Corresponding Configs
data augmentation	no aug	rotate & flip
Dnet input	radial distance	radial distance & height
Dnet structure	last1 pruning	last2 pruning	last3 pruning	last4 pruning	Dnet relu	4convs output
Gnet structure	width128 depth4	width512 depth4	width256 depth3	width256 depth5	Gnet relu
point sample	on:off=1:2	on:off=2:3
positional encoding	no encoding	incF level10	incT level5	incT level15
sample strategy	nearest
scale size	scale 2	scale 4	scale 8	scale 16	scale 32
shape size	shape 128	shape 512
SSC	SSC option A	SSC option B