search

cvg / nice-slam

[CVPR'22] NICE-SLAM: Neural Implicit Scalable Encoding for SLAM
https://pengsongyou.github.io/nice-slam
Apache License 2.0
1.39k stars 192 forks source link
3d-reconstruction deep-learning implicit-functions localization neural-fields neural-implicit-representations scalable slam

readme

NICE-SLAM: Neural Implicit Scalable Encoding for SLAM

Zihan Zhu* · Songyou Peng* · Viktor Larsson · Weiwei Xu · Hujun Bao
Zhaopeng Cui · Martin R. Oswald · Marc Pollefeys

(* Equal Contribution)

CVPR 2022

Paper | Video | Project Page

Logo

NICE-SLAM produces accurate dense geometry and camera tracking on large-scale indoor scenes.

(The black / red lines are the ground truth / predicted camera trajectory)



Table of Contents
  1. Installation
  2. Visualization
  3. Demo
  4. Run
  5. iMAP*
  6. Evaluation
  7. Acknowledgement
  8. Citation
  9. Contact

Installation

First you have to make sure that you have all dependencies in place. The simplest way to do so, is to use anaconda.

You can create an anaconda environment called nice-slam. For linux, you need to install libopenexr-dev before creating the environment.

sudo apt-get install libopenexr-dev

conda env create -f environment.yaml
conda activate nice-slam

Visualizing NICE-SLAM Results

We provide the results of NICE-SLAM ready for download. You can run our interactive visualizer as following.

Self-captured Apartment

To visualize our results on the self-captured apartment, as shown in the teaser:

bash scripts/download_vis_apartment.sh
python visualizer.py configs/Apartment/apartment.yaml --output output/vis/Apartment

Note for users from China: If you encounter slow speed in downloading, check in all the scripts/download_*.sh scripts, where we also provide the 和彩云 links for you to download manually.

ScanNet

bash scripts/download_vis_scene0000.sh
python visualizer.py configs/ScanNet/scene0000.yaml --output output/vis/scannet/scans/scene0000_00

You can find the results of NICE-SLAM on other scenes in ScanNet here.

Replica

bash scripts/download_vis_room1.sh
python visualizer.py configs/Replica/room1.yaml --output output/vis/Replica/room1

[Directory structure of ScanNet (click to expand)] DATAROOT is `./Datasets` by default. If a sequence (`sceneXXXX_XX`) is stored in other places, please change the `input_folder` path in the config file or in the command line. ``` DATAROOT └── scannet └── scans └── scene0000_00 └── frames ├── color │ ├── 0.jpg │ ├── 1.jpg │ ├── ... │ └── ... ├── depth │ ├── 0.png │ ├── 1.png │ ├── ... │ └── ... ├── intrinsic └── pose ├── 0.txt ├── 1.txt ├── ... └── ... ``` Once the data is downloaded and set up properly, you can run NICE-SLAM: ```bash python -W ignore run.py configs/ScanNet/scene0000.yaml ``` ### Replica Download the data as below and the data is saved into the `./Datasets/Replica` folder. Note that the Replica data is generated by the authors of iMAP, so please cite iMAP if you use the data. ```bash bash scripts/download_replica.sh ``` and you can run NICE-SLAM: ```bash python -W ignore run.py configs/Replica/room0.yaml ``` The mesh for evaluation is saved as `$OUTPUT_FOLDER/mesh/final_mesh_eval_rec.ply`, where the unseen regions are culled using all frames. ### TUM RGB-D Download the data as below and the data is saved into the `./Datasets/TUM-RGBD` folder ```bash bash scripts/download_tum.sh ``` Now run NICE-SLAM: ```bash python -W ignore run.py configs/TUM_RGBD/freiburg1_desk.yaml ``` ### Co-Fusion First, download the dataset. This script should download and unpack the data automatically into the `./Datasets/CoFusion` folder. ```bash bash scripts/download_cofusion.sh ``` Run NICE-SLAM: ```bash python -W ignore run.py configs/CoFusion/room4.yaml ``` ### Use your own RGB-D sequence from Kinect Azure
[Details (click to expand)] 1. Please first follow this [guide](http://www.open3d.org/docs/release/tutorial/sensor/azure_kinect.html#install-the-azure-kinect-sdk) to record a sequence and extract aligned color and depth images. (Remember to use `--align_depth_to_color` for `azure_kinect_recorder.py`) DATAROOT is `./Datasets` in default, if a sequence (`sceneXX`) is stored in other places, please change the "input_folder" path in the config file or in the command line. ``` DATAROOT └── Own └── scene0 ├── color │ ├── 00000.jpg │ ├── 00001.jpg │ ├── 00002.jpg │ ├── ... │ └── ... ├── config.json ├── depth │ ├── 00000.png │ ├── 00001.png │ ├── 00002.png │ ├── ... │ └── ... └── intrinsic.json ``` 2. Prepare `.yaml` file based on the `configs/Own/sample.yaml`. Change the camera intrinsics in the config file based on `intrinsic.json`. You can also get the intrinsics of the depth camera via other tools such as MATLAB. 3. Specify the bound of the scene. If no ground truth camera pose is given, we construct world coordinates on the first frame. The X-axis is from left to right, Y-axis is from down to up, Z-axis is from front to back. 4. Change the `input_folder` path and/or the `output` path in the config file or the command line. 5. Run NICE-SLAM. ```bash python -W ignore run.py configs/Own/sample.yaml ``` **(Optional but highly Recommended)** If you don't want to specify the bound of the scene or manually change the config file. You can first run the Redwood tool in [Open3D](http://www.open3d.org/) and then run NICE-SLAM. Here we provide steps for the whole pipeline, beginning from recording Azure Kinect videos. (Ubuntu 18.04 and above is recommended.) 1. Download the Open3D repository. ```bash bash scripts/download_open3d.sh ``` 2. Record and extract frames. ```bash # specify scene ID sceneid=0 cd 3rdparty/Open3D-0.13.0/examples/python/reconstruction_system/ # record and save to .mkv file python sensors/azure_kinect_recorder.py --align_depth_to_color --output scene$sceneid.mkv # extract frames python sensors/azure_kinect_mkv_reader.py --input scene$sceneid.mkv --output dataset/scene$sceneid ``` 3. Run reconstruction. ```bash python run_system.py dataset/scene$sceneid/config.json --make --register --refine --integrate # back to main folder cd ../../../../../ ``` 4. Prepare the config file. ```bash python src/tools/prep_own_data.py --scene_folder 3rdparty/Open3D-0.13.0/examples/python/reconstruction_system/dataset/scene$sceneid --ouput_config configs/Own/scene$sceneid.yaml ``` 5. Run NICE-SLAM. ```bash python -W ignore run.py configs/Own/scene$sceneid.yaml ```
## iMAP* We also provide our re-implementation of iMAP (iMAP*) for use. If you use the code, please cite both the original iMAP paper and NICE-SLAM. ### Usage iMAP* shares a majority part of the code with NICE-SLAM. To run iMAP*, simply use `*_imap.yaml` in the config file and also add the argument `--imap` in the command line. For example, to run iMAP* on Replica room0: ```bash python -W ignore run.py configs/Replica/room0_imap.yaml --imap ``` To use our interactive visualizer: ```bash python visualizer.py configs/Replica/room0_imap.yaml --imap ``` To evaluate ATE: ```bash python src/tools/eval_ate.py configs/Replica/room0_imap.yaml --imap ```
[Differences between iMAP* and the original iMAP (click to expand)] #### Keyframe pose optimization during mapping We do not optimize the selected keyframes' poses for iMAP*, because optimizing them usually leads to worse performance. One possible reason is that since their keyframes are selected globally, and many of them do not have overlapping regions especially when the scene gets larger. Overlap is a prerequisite for bundle adjustment (BA). For NICE-SLAM, we only select overlapping keyframes within a small window (local BA), which works well in all scenes. You can still turn on the keyframe pose optimization during mapping for iMAP* by enabling `BA` in the config file. #### Active sampling We disable the active sampling in iMAP*, because in our experiments we observe that it does not help to improve the performance while brings additional computational overhead. For the image active sampling, in each iteration the original iMAP uniformly samples 200 pixels in the entire image. Next, they divide this image into an 8x8 grid and calculate the probability distribution from the rendering losses. This means that if the resolution of an image is 1200x680 (Replica), only around 3 pixels are sampled to calculate the distribution for a 150x85 grid patch. This is not too much different from simple uniform sampling. Therefore, during mapping we use the same pixel sampling strategy as NICE-SLAM for iMAP*: uniform sampling, but even 4x more pixels than reported in the iMAP paper. For the keyframe active sampling, the original iMAP requires rendering depth and color images for all keyframes to get the loss distribution, which is expensive and we again did not find it very helpful. Instead, as done in NICE-SLAM, iMAP* randomly samples keyframes from the keyframe list. We also let iMAP* optimize for 4x more iterations than NICE-SLAM, but their performance is still inferior. #### Keyframe selection For fair comparison, we use the same keyframe selection method in iMAP* as in NICE-SLAM: add one keyframe to the keyframe list every 50 frames.
## Evaluation ### Average Trajectory Error To evaluate the average trajectory error. Run the command below with the corresponding config file: ```bash python src/tools/eval_ate.py configs/Replica/room0.yaml ``` ### Reconstruction Error To evaluate the reconstruction error, first download the ground truth Replica meshes where unseen region have been culled. ```bash bash scripts/download_cull_replica_mesh.sh ``` Then run the command below (same for NICE-SLAM and iMAP*). The 2D metric requires rendering of 1000 depth images, which will take some time (~9 minutes). Use `-2d` to enable 2D metric. Use `-3d` to enable 3D metric. ```bash # assign any output_folder and gt mesh you like, here is just an example OUTPUT_FOLDER=output/Replica/room0 GT_MESH=cull_replica_mesh/room0.ply python src/tools/eval_recon.py --rec_mesh $OUTPUT_FOLDER/mesh/final_mesh_eval_rec.ply --gt_mesh $GT_MESH -2d -3d ``` We also provide code to cull the mesh given camera poses. Here we take culling of ground truth mesh of Replica room0 as an example. ```bash python src/tools/cull_mesh.py --input_mesh Datasets/Replica/room0_mesh.ply --traj Datasets/Replica/room0/traj.txt --output_mesh cull_replica_mesh/room0.ply ```
[For iMAP* evaluation (click to expand)] As discussed in many recent papers, e.g. UNISURF/VolSDF/NeuS, manual thresholding the volume density during marching cubes might be needed. Moreover, we find out there exist scaling differences, possibly because of the reason discussed in [NeuS](https://arxiv.org/abs/2106.10689). Therefore, ICP with scale is needed. You can use the [ICP tool](https://www.cloudcompare.org/doc/wiki/index.php?title=ICP) in [CloudCompare](https://www.danielgm.net/cc/) with default configuration with scaling enabled.
## Acknowledgement We adapted some codes from some awesome repositories including [convolutional_occupancy_networks](https://github.com/autonomousvision/convolutional_occupancy_networks), [nerf-pytorch](https://github.com/yenchenlin/nerf-pytorch), [lietorch](https://github.com/princeton-vl/lietorch), and [DIST-Renderer](https://github.com/B1ueber2y/DIST-Renderer). Thanks for making codes public available. We also thank [Edgar Sucar](https://edgarsucar.github.io/) for allowing us to make the Replica Dataset available. ## Citation If you find our code or paper useful, please cite ```bibtex @inproceedings{Zhu2022CVPR, author = {Zhu, Zihan and Peng, Songyou and Larsson, Viktor and Xu, Weiwei and Bao, Hujun and Cui, Zhaopeng and Oswald, Martin R. and Pollefeys, Marc}, title = {NICE-SLAM: Neural Implicit Scalable Encoding for SLAM}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, year = {2022} } ``` ## Contact Contact [Zihan Zhu](mailto:zhuzihan2000@gmail.com) and [Songyou Peng](mailto:songyou.pp@gmail.com) for questions, comments and reporting bugs.