Home

Awesome

Point-NeRF: Point-based Neural Radiance Fields (CVPR 2022 Oral 🀩)

<img src="./images/Adobe-Logos.png" width=120px /><img src="images/USC-Logos.png" width=120px />

Project Sites | Paper | Primary contact: Qiangeng Xu

Point-NeRF uses neural 3D point clouds, with associated neural features, to model a radiance field. Point-NeRF can be rendered efficiently by aggregating neural point features near scene surfaces, in a ray marching-based rendering pipeline. Moreover, Point-NeRF can be initialized via direct inference of a pre-trained deep network to produce a neural point cloud; this point cloud can be finetuned to surpass the visual quality of NeRF with 30X faster training time. Point-NeRF can be combined with other 3D reconstruction methods and handles the errors and outliers in such methods via a novel pruning and growing mechanism.

<!-- <img src="./images/pipeline.png" /> -->

CVPR 2022 Oral Presentation

Reference

Please cite our paper if you are interested
<strong>Point-NeRF: Point-based Neural Radiance Fields</strong>. Β Β Β 

@inproceedings{xu2022point,
  title={Point-nerf: Point-based neural radiance fields},
  author={Xu, Qiangeng and Xu, Zexiang and Philip, Julien and Bi, Sai and Shu, Zhixin and Sunkavalli, Kalyan and Neumann, Ulrich},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={5438--5448},
  year={2022}
}

Updates

  1. To replace pycuda, we have implemented the pytorch cuda functions when using world coordinates to group neural points. Simply set wcoord_query=-1 in your configuration file if the original setting is wcoord_query=1 (see dev_scripts/w_n360/chair_cuda.sh).
  2. We have received constructive feedbacks that when Point-NeRF use MVSNet to reconstruct point cloud, the point fusion after depth estimation by MVSNet will use the alpha channel information in the NeRF-Synthetic Dataset. It is due to the fact that MVSNet cannot handle background very well. To improve the fairness, we include new training scripts and results of PointNeRF + MVSNet when using background color for filtering. The results (see below) are similar to the ones that are previously reported.
ChairDrumsLegoMicMaterialsShipHotdogFicusAvg
PSNR35.6026.0435.2735.9129.6530.6137.3435.6133.25
SSIM0.9910.9540.9890.9940.9710.9380.9910.9920.978
LPIPSVgg0.0230.0780.0210.0140.0710.1290.0360.0250.050
LPIPSAlex0.0100.0550.0100.0070.0410.0760.0160.0110.028

This issue only affacts situations when Point-NeRF uses MVSNet on NeRF-Synthetic Dataset. The Colmap results and results on other datasets are not impacted.
An even more reasonable reconstruction approach should exclude using the knowledge of background color or other point filtering. Therefore, we suggest users to combine PointNeRF with more powerful MVS models, such as TransMVS.

Overal Instruction

  1. Please first install the libraries as below and download/prepare the datasets as instructed.
  2. Point Initialization: Download pre-trained MVSNet as below and train the feature extraction from scratch or directly download the pre-trained models. (Obtain 'MVSNet' and 'init' folder in checkpoints folder)
  3. Per-scene Optimization: Download pre-trained models or optimize from scratch as instructed.

For nerfsynthetic, colmap_nerfsynthetic, tanks&temples, scannet and dtu,
We provide all the checkpoint_files google drive | baidu wangpan; all the images and scores of the test results google drive | baidu wangpan; and video results google drive | baidu wangpan;

We also share the visual results of npbg, nsvf and ibrnet on the Nerf Synthetic dataset generated by our machine google drive; Please cite their papers accordingly if interested.

Installation

Requirements

All the codes are tested in the following environment:

Install

Install the dependent libraries as follows:

pip install torch==1.8.1+cu102 h5py
pip install imageio scikit-image

We develope our code with pytorch1.8.1, pycuda2021.1, and torch_scatter 2.0.8

Data Preparation

The layout should looks like this, we provide all data folder here: google_drive, except for scannet (We take it down on Mar 8th 2023, to respect Scannet's policy, please go to Scannet's official website for data.)

pointnerf
β”œβ”€β”€ data_src
β”‚   β”œβ”€β”€ dtu
    β”‚   β”‚   │──Cameras
    β”‚   β”‚   │──Depths
    β”‚   β”‚   │──Depths_raw
    β”‚   β”‚   │──Rectified
    β”œβ”€β”€ nerf
    β”‚   β”‚   │──nerf_synthetic
    β”‚   β”‚   │──nerf_synthetic_colmap
    β”œβ”€β”€ TanksAndTemple
    β”œβ”€β”€ scannet
    β”‚   β”‚   │──scans 
    |   β”‚   β”‚   │──scene0101_04
    |   β”‚   β”‚   │──scene0241_01

Or you can download using the official links as follows:

DTU:

Download the preprocessed DTU training data and Depth_raw from original MVSNet repo and unzip.

NeRF Synthetic

Download nerf_synthetic.zip from here under ``data_src/nerf/''

Tanks & Temples

Follow Neural Sparse Voxel Fields and download Tanks&Temples | download (.zip) | 0_* (training) 1_* (testing) as: ``data_src/TanksAndTemple/''

ScanNet

Download and extract ScanNet by following the instructions provided at http://www.scan-net.org/. The detailed steps including:

Point Initialization / Generalization:

Β  Download pre-trained MVSNet checkpoints:

We trained MVSNet on DTU. You can Download ''MVSNet'' directory from google drive and place them under '''checkpoints/'''

Β  Train 2D feature extraction and point representation

Β  Directly use our trained checkpoints files:

Download ''init'' directory from google drive. and place them under '''checkpoints/'''

Β  Or train from scratch:

Train for point features of 63 channels (as in paper)

bash dev_scripts/ete/dtu_dgt_d012_img0123_conf_color_dir_agg2.sh

Train for point features of 32 channels (better for per-scene optimization)

bash dev_scripts/ete/dtu_dgt_d012_img0123_conf_agg2_32_dirclr20.sh

After the training, you should pick a checkpoint and rename it to best checkpoint, e.g.:

cp checkpoints/dtu_dgt_d012_img0123_conf_color_dir_agg2/250000_net_ray_marching.pth  checkpoints/dtu_dgt_d012_img0123_conf_color_dir_agg2/best_net_ray_marching.pth

cp checkpoints/dtu_dgt_d012_img0123_conf_color_dir_agg2/250000_net_mvs.pth  checkpoints/dtu_dgt_d012_img0123_conf_color_dir_agg2/best_net_mvs.pth

Β  Test feed forward inference on dtu scenes

These scenes that are selected by MVSNeRF, please also refer their code to understand the metrics calculation.

bash dev_scripts/dtu_test_inf/inftest_scan1.sh
bash dev_scripts/dtu_test_inf/inftest_scan8.sh
bash dev_scripts/dtu_test_inf/inftest_scan21.sh
bash dev_scripts/dtu_test_inf/inftest_scan103.sh
bash dev_scripts/dtu_test_inf/inftest_scan114.sh

Per-scene Optimization:

<img src="https://github.com/Xharlie/xharlie.github.io/raw/master/projects/project_sites/pointnerf/vid/ficus.gif" width="45%" /><img src="https://github.com/Xharlie/xharlie.github.io/raw/master/projects/project_sites/pointnerf/vid/scene101.gif" width="50%" /> <img src="https://github.com/Xharlie/xharlie.github.io/raw/master/projects/project_sites/pointnerf/vid/truck.gif" width="70%" />

(Please visit the project sites to see the original videos of above scenes, which have quality loss when being converted to gif files here.)

Download per-scene optimized Point-NeRFs

You can skip training and download the folders of ''nerfsynth'', ''tanksntemples'' and ''scannet'' here google drive, and place them in ''checkpoints/''.

pointnerf
β”œβ”€β”€ checkpoints
β”‚   β”œβ”€β”€ init
    β”œβ”€β”€ MVSNet
    β”œβ”€β”€ nerfsynth
    β”œβ”€β”€ col_nerfsynth
    β”œβ”€β”€ scannet
    β”œβ”€β”€ tanksntemples

In each scene, we provide initialized point features and network weights ''0_net_ray_marching.pth'', points and weights at 20K steps ''20000_net_ray_marching.pth'' and 200K steps ''200000_net_ray_marching.pth''

Test the per-scene optimized Point-NeRFs

NeRF Synthetics

<details> <summary>test scripts</summary>
    bash dev_scripts/w_n360/chair_test.sh
    bash dev_scripts/w_n360/drums_test.sh
    bash dev_scripts/w_n360/ficus_test.sh
    bash dev_scripts/w_n360/hotdog_test.sh
    bash dev_scripts/w_n360/lego_test.sh
    bash dev_scripts/w_n360/materials_test.sh
    bash dev_scripts/w_n360/mic_test.sh
    bash dev_scripts/w_n360/ship_test.sh
</details>

ScanNet

<details> <summary>test scripts</summary>
    bash dev_scripts/w_scannet_etf/scane101_test.sh
    bash dev_scripts/w_scannet_etf/scane241_test.sh
</details>

Tanks & Temples

<details> <summary>test scripts</summary>
    bash dev_scripts/w_tt_ft/barn_test.sh
    bash dev_scripts/w_tt_ft/caterpillar_test.sh
    bash dev_scripts/w_tt_ft/family_test.sh
    bash dev_scripts/w_tt_ft/ignatius_test.sh
    bash dev_scripts/w_tt_ft/truck_test.sh
</details>

Per-scene optimize from scatch

Make sure the ''checkpoints'' folder has ''init'' and ''MVSNet''. The training scripts will start to do initialization if there is no ''.pth'' files in a scene folder. It will start from the last ''.pth'' files until reach the iteration of ''maximum_step''.

NeRF Synthetics using MVSNet (w/ alpha channel filtering during point cloud reconstruction and pycuda)

<details> <summary>train scripts</summary>
    bash dev_scripts/w_n360/chair.sh
    bash dev_scripts/w_n360/drums.sh
    bash dev_scripts/w_n360/ficus.sh
    bash dev_scripts/w_n360/hotdog.sh
    bash dev_scripts/w_n360/lego.sh
    bash dev_scripts/w_n360/materials.sh
    bash dev_scripts/w_n360/mic.sh
    bash dev_scripts/w_n360/ship.sh
</details>

NeRF Synthetics using MVSNet (w/ background color filtering during point cloud reconstruction and pytorch cuda)

<details> <summary>train scripts</summary>
    bash dev_scripts/w_n360/chair_cuda.sh
    bash dev_scripts/w_n360/drums_cuda.sh
    bash dev_scripts/w_n360/ficus_cuda.sh
    bash dev_scripts/w_n360/hotdog_cuda.sh
    bash dev_scripts/w_n360/lego_cuda.sh
    bash dev_scripts/w_n360/materials_cuda.sh
    bash dev_scripts/w_n360/mic_cuda.sh
    bash dev_scripts/w_n360/ship_cuda.sh
</details>

NeRF Synthetics using COLMAP points

Please download the COLMAP data (see above). If there is {maximum_step}.pth checkpoint files in the path, the scripts below will also run test.

<details> <summary>train scripts</summary>
    bash dev_scripts/w_colmap_n360/col_chair.sh
    bash dev_scripts/w_colmap_n360/col_drums.sh
    bash dev_scripts/w_colmap_n360/col_ficus.sh
    bash dev_scripts/w_colmap_n360/col_hotdog.sh
    bash dev_scripts/w_colmap_n360/col_lego.sh
    bash dev_scripts/w_colmap_n360/col_materials.sh
    bash dev_scripts/w_colmap_n360/col_mic.sh
    bash dev_scripts/w_colmap_n360/col_ship.sh
</details>

ScanNet

<details> <summary>train scripts</summary>
    bash dev_scripts/w_scannet_etf/scene101.sh
    bash dev_scripts/w_scannet_etf/scene241.sh
</details>

Tanks & Temples

<details> <summary>train scripts</summary>
    bash dev_scripts/w_tt_ft/barn.sh
    bash dev_scripts/w_tt_ft/caterpillar.sh
    bash dev_scripts/w_tt_ft/family.sh
    bash dev_scripts/w_tt_ft/ignatius.sh
    bash dev_scripts/w_tt_ft/truck.sh
</details>

Acknowledgement

Our repo is developed based on MVSNet, NeRF, MVSNeRF, and NSVF.

Please also consider citing the corresponding papers.

The project is conducted collaboratively between Adobe Research and University of Southern California.

LICENSE

The repo is licensed under Creative Commons Attribution-NonCommercial-ShareAlike 2.0, and is restricted to academic use only. See LICENSE.