Qianyi Wu · Kaisiyuan Wang · Kejie Li · Jianmin Zheng · Jianfei Cai

TL; DR: We propose an occlusion-aware opacity rendering formulation to better use the instance mask supervision. Together with an object-distinction regularization term, the proposed ObjectSDF++ produces more accurate surface reconstruction at both scene and object levels.

This code has been tested on Ubuntu 22.02 with torch 2.0 & CUDA 11.7 on a RTX 3090. Clone the repository and create an anaconda environment named objsdf

git clone https://github.com/QianyiWu/objectsdf_plus.git
cd objectsdf_plus

conda create -y -n objsdf python=3.9
conda activate object

pip install -r requirements.txt

The hash encoder will be compiled on the fly when running the code.

For downloading the preprocessed data, run the following script. The data for the Replica and ScanNet is adapted from MonoSDF, vMAP.

bash scripts/download_dataset.sh

Run the following command to train ObjectSDF++:

cd ./code
CUDA_VISIBLE_DEVICES=0 torchrun --nproc_per_node=1 --nnodes=1 --node_rank=0 training/exp_runner.py --conf CONFIG  --scan_id SCAN_ID

where CONFIG is the config file in code/confs, and SCAN_ID is the id of the scene to reconstruct.

We provide example commands for training Replica dataset as follows:

# Replica scan 1 (room0)
CUDA_VISIBLE_DEVICES=0 torchrun --nproc_per_node=1 --nnodes=1 --node_rank=0 training/exp_runner.py --conf confs/replica_objsdfplus.conf --scan_id 1

# ScanNet scan 1 (scene_0050_00)
CUDA_VISIBLE_DEVICES=0 torchrun --nproc_per_node=1 --nnodes=1 --node_rank=0 training/exp_runner.py --conf confs/scannet_objsdfplus.conf --scan_id 1

The intermediate results and checkpoints will be saved in exps folder.

Evaluate one scene (take scan 1 room0 for example)

cd replica_eval
python evaluate_single_scene.py --input_mesh replica_scan1_mesh.ply --scan_id 1 --output_dir replica_scan1

We also provided scripts for evaluating all Replica scenes and objects:

cd replica_eval
python evaluate.py  # scene-level evaluation
python evaluate_3D_obj.py   # object-level evaluation

please check the script for more details. For obtaining the object groundtruth, you can refer to here for more details.

cd scannet_eval
python evaluate.py

please check the script for more details.

This project is built upon MonoSDF. The monocular depth and normal images are obtained by Omnidata. The evaluation of object reconstruction is inspired by vMAP. Cuda implementation of Multi-Resolution hash encoding is heavily based on torch-ngp. Kudos to these researchers.

If you find our code or paper useful, please cite the series of ObjectSDF works.

@inproceedings{wu2022object,
  title      = {Object-compositional neural implicit surfaces},
  author     = {Wu, Qianyi and Liu, Xian and Chen, Yuedong and Li, Kejie and Zheng, Chuanxia and Cai, Jianfei and Zheng, Jianmin},
  booktitle  = {European Conference on Computer Vision},
  year       = {2022},
}

@inproceedings{wu2023objsdfplus,
  author    = {Wu, Qianyi and Wang, Kaisiyuan and Li, Kejie and Zheng, Jianmin and Cai, Jianfei},
  title     = {ObjectSDF++: Improved Object-Compositional Neural Implicit Surfaces},
  booktitle = {ICCV},
  year      = {2023},
}