Panoptic NeRF: 3D-to-2D Label Transfer for Panoptic Urban Scene Segmentation

Xiao Fu¹*  Shangzhan Zhang¹*  Tianrun Chen¹  Yichong Lu¹  Lanyun Zhu²  Xiaowei Zhou²  Andreas Geiger³  Yiyi Liao^1†

¹Zhejiang University  ²Singapore University of Technology and Design  ³University of Tübingen and MPI-IS

*equal contribution  ⁺corresponding author

3DV 2022

[Project Page] • [Paper] • [Dataset] • [Video] • [Poster]

Updates

• [08/2022] Paper accepted to 3DV 2022.
• [04/2022] Code released.
• [03/2022] Paper uploaded to arXiv.

Installation

1. Create a virtual environment via conda. This environment is built on RTX 3090 and can be modified manually.

   conda env create -f environment.yml
   conda activate panopticnerf
   

Data Preparation

1. We evaluate our model on KITTI-360. Here we show the structure of a test dataset as follow. You can download it from here and then put it into $ROOT. In the datasets, we additionally provide some test files on different scenes.

   ├── KITTI-360
     ├── 2013_05_28_drive_0000_sync
       ├── image_00
       ├── image_01
     ├── bbx_intersection
       ├── *_00.npz
       ├── *_01.npz
     ├── calibration
       ├── calib_cam_to_pose.txt
       ├── perspective.txt
     ├── data_3d_bboxes
     ├── data_poses
       ├── cam0_to_world.txt
       ├── poses.txt
     ├── pspnet
     ├── sgm
     ├── visible_id
   

   file	Intro
   image_00/01	stereo RGB images
   pspnet	2D pseudo ground truth
   sgm	weak stereo depth supervision
   visible_id	per-frame bounding primitive IDs
   data_poses	system poses in a global Euclidean coordinate
   calibration	extrinsics and intrinsics of the perspective cameras
   bbx_intersection	ray-mesh intersections, containing depths between hitting points and camera origin, semantic label IDs and bounding primitive IDs

2. Generate ray-mesh intersections (bbx_intersection/*.npz). The red dots and blue dots indicate where the rays hit into and out of the meshes, respectively. For the given test scene, START=3353, NUM=64.

   # image_00
   python mesh_intersection.py intersection_start_frame ${START} intersection_frames ${NUM} use_stereo False
   # image_01
   python mesh_intersection.py intersection_start_frame ${START} intersection_frames ${NUM} use_stereo True
   

3. Evaluate the origin of a scene (center_pose) and the distance from the origin to the furthest bounding primitive (dist_min). Then accordingly modify the .yaml file.

   python recenter_pose.py recenter_start_frame ${START} recenter_frames ${NUM}
   

Training and Visualization

1. We provide the training code. Replace resume False with resume True to load the pretained model.

   python train_net.py --cfg_file configs/panopticnerf_test.yaml pretrain nerf gpus '1,' use_stereo True use_pspnet True use_depth True pseudo_filter True weight_th 0.05 resume False
   

2. Render semantic map, panoptic map and depth map in a single forward pass, which takes around 16s per-frame on a single 3090 GPU. Please make sure to maximize the GPU memory utilization by increasing the size of the chunk to reduce inference time. Replace use_stereo False with use_stereo True to render the right views.

   python run.py --type visualize --cfg_file configs/panopticnerf_test.yaml use_stereo False
   

   

3. Visualize novel view appearance & label synthesis. Before rendering, select a frame and generate corresponding ray-mesh intersections with respect to its novel spiral poses, e.g. SPIRAL_FRAME=3400, NUM=32.

   python mesh_intersection_spiral_trajectory.py intersection_spiral_frame ${SPIRAL_FRAME} intersection_frames ${NUM} use_stereo False
   

   Then render results of the spiral trajectory. Feel free to change codes for rendering from arbitrary poses.

   python run.py --type visualize --cfg_file configs/panopticnerf_test_spiral.yaml spiral_frame ${SPIRAL_FRAME} spiral_frame_num ${NUM} use_stereo False
   

   

Evaluation

├── KITTI-360
  ├── gt_2d_semantics
  ├── gt_2d_panoptics
  ├── lidar_depth

1. Download the released pretrained model and put it to $ROOT/data/trained_model/panopticnerf/panopticnerf_test/latest.pth.

2. We provide some semantic & panoptic GTs and LiDAR point clouds for evaluation. The details of evaluation metrics can be found in the paper.

3. Eval mean intersection-over-union (mIoU)

python run.py --type eval_miou --cfg_file configs/panopticnerf_test.yaml use_stereo False

4. Eval panoptic quality (PQ)

sh eval_pq_test.sh

5. Eval depth with 0-100m LiDAR point clouds, where the far depth can be adjusted to evaluate the closer scene.

python run.py --type eval_depth --cfg_file configs/panopticnerf_test.yaml use_stereo False max_depth 100.

6. Eval Multi-view Consistency (MC)

python eval_consistency.py --cfg_file configs/panopticnerf_test.yaml use_stereo False consistency_thres 0.1

Citation

@inproceedings{fu2022panoptic,
  title={Panoptic NeRF: 3D-to-2D Label Transfer for Panoptic Urban Scene Segmentation},
  author={Fu, Xiao and Zhang, Shangzhan and Chen, Tianrun and Lu, Yichong and Zhu, Lanyun and Zhou, Xiaowei and Geiger, Andreas and Liao, Yiyi},
  booktitle = {International Conference on 3D Vision (3DV)},
  year = {2022}
}

Copyright © 2022, Zhejiang University. All rights reserved. We favor any positive inquiry, please contact lemonaddie0909@zju.edu.cn.