Skip to content

The implementation of "In-Place Scene Labelling and Understanding with Implicit Scene Representation" [ICCV 2021].


Notifications You must be signed in to change notification settings


Folders and files

Last commit message
Last commit date

Latest commit



12 Commits

Repository files navigation

Semantic-NeRF: Semantic Neural Radiance Fields

In-Place Scene Labelling and Understanding with Implicit Scene Representation
Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, Andrew J. Davison,
Dyson Robotics Laboratory at Imperial College
Published in ICCV 2021 (Oral Presentation)

We build upon neural radiance fields to create a scene-specific implicit 3D semantic representation, Semantic-NeRF.

Latest Updates.

  • Release of Replica Data Generation Codes. We have provided data generation scripts for Replica sequences at SSR/data_generation folder. Thanks Xin of vMAP for cleaning up.
  • Instance Label Maps Available. We have also provided corresponding instance label maps of pre-rendered Replica sequences in dropbox as a zip file

Getting Started

For flawless reproduction of our results, the Ubuntu OS 20.04 is recommended. The models have been tested using Python 3.7, Pytorch 1.6.0, CUDA10.1. Higher versions should also perform similarly.


Main python dependencies are listed below:

  • Python >=3.7
  • torch>=1.6.0 (integrate searchsorted API, otherwise need to use the third party implementation SearchSorted )
  • cudatoolkit>=10.1

Following packages are used for 3D mesh reconstruction:

  • trimesh==3.9.9
  • open3d==0.12.0

With Anaconda, you can simply create a virtual environment and install dependencies with CONDA by:

  • conda create -n semantic_nerf python=3.7
  • conda activate semantic_nerf
  • pip install -r requirements.txt


We mainly use Replica and ScanNet datasets for experiments, where we train a new Semantic-NeRF model on each 3D scene. Other similar indoor datasets with colour images, semantic labels and poses can also be used.

We also provide pre-rendered Replica data that can be directly used by Semantic-NeRF.

Running code

After cloning the codes, we can start to run Semantic-NeRF in the root directory of the repository.

Semantic-NeRF training

For standard Semantic-NeRF training with full dense semantic supervision. You can simply run following command with a chosen config file specifying data directory and hyper-params.

python3 --config_file /SSR/configs/SSR_room0_config.yaml

Different working modes and set-ups can be chosen via commands:

Semantic View Synthesis with Sparse Labels:

python3 --sparse_views --sparse_ratio 0.6

Sparse ratio here is the portion of dropped frames in the training sequence.

Pixel-wise Denoising Task:

python3 --pixel_denoising --pixel_noise_ratio 0.5

We could also use a sparse set of frames along with denoising task:

python3 --pixel_denoising --pixel_noise_ratio 0.5 --sparse_views --sparse_ratio 0.6

Region-wise Denoising task (For Replica Room2):

python3 --region_denoising --region_noise_ratio 0.3

The argument uniform_flip corresponds to the two modes of "Even/Sort"in region-wise denoising task.

Super-Resolution Task:

For super-resolution with dense labels, please run

python3 --super_resolution --sr_factor 8 --dense_sr

For super-resolution with sparse labels, please run

python3 --super_resolution --sr_factor 8

Label Propagation Task:

For label propagation task with single-click seed regions, please run

python3 --label_propagation --partial_perc 0

In order to improve reproducibility, for denoising and label-propagation tasks, we can also include --visualise_save and --load_saved to save/load randomly generated labels.

3D Reconstruction of Replica Scenes

We also provide codes for extracting 3D semantic mesh from a trained Seamntic-NeRF model.

python3 SSR/ --sem --mesh_dir PATH_TO_MESH --mesh_dir PATH_TO_MESH  --training_data_dir PATH_TO_TRAINING_DATA --save_dir PATH_TO_SAVE_DIR

For more demos and qualitative results, please check our project page and video.


Thanks nerf, nerf-pytorch and nerf_pl for providing nice and inspiring implementations of NeRF. Thank Atlas for scripts in processing ScanNet dataset.


If you found this code/work to be useful in your own research, please consider citing the following:

  title={In-Place Scene Labelling and Understanding with Implicit Scene Representation},
  author={Shuaifeng Zhi and Tristan Laidlow and Stefan Leutenegger and Andrew J. Davison},


If you have any questions, please contact or


The implementation of "In-Place Scene Labelling and Understanding with Implicit Scene Representation" [ICCV 2021].







No releases published