idn-solver

Paper | Project Page

This repository contains the code release of our ICCV 2021 paper:

A Confidence-based Iterative Solver of Depths and Surface Normals for Deep Multi-view Stereo

Wang Zhao*, Shaohui Liu*, Yi Wei, Hengkai Guo, Yong-Jin Liu

Installation

We recommend to use conda to setup a specified environment. Run

conda env create -f environment.yml

Test on a sequence

First download the pretrained model from here and put it under ./pretrain/ folder.

Prepare the sequence data with color images, camera poses (4x4 cam2world transformation) and intrinsics. The sequence data structure should be like:

sequence_name
  | color
      | 00000.jpg
  | pose
      | 00000.txt
  | K.txt

Run the following command to get the outputs:

python infer_folder.py --seq_dir /path/to/the/sequence/data --output_dir /path/to/save/outputs --config ./configs/test_folder.yaml

Tune the "reference gap" parameter to make sure there are sufficient overlaps and camera translations within an image pair. For ScanNet-like sequence, we recommend to use reference_gap of 20.

Test on ScanNet

Prepare ScanNet test split data

Download the ScanNet test split data from the official site and pre-process the data using:

python ./data/preprocess.py --data_dir /path/to/scannet/test/split/ --output_dir /path/to/save/pre-processed/scannet/test/data

This includes 1. resize the color images to 480x640 resolution 2. sample the data with interval of 20

Run evaluation

python eval_scannet.py --data_dir /path/to/processed/scannet/test/split/ --config ./configs/test_scannet.yaml

Train

Prepare ScanNet training data

We use the pre-processed ScanNet data from NAS, you could download the data using this link. The data structure is like:

scannet
  | scannet_nas
    | train
      | scene0000_00
          | color
            | 0000.jpg
          | pose
            | 0000.txt
          | depth
            | 0000.npy
          | intrinsic
          | normal
            | 0000_normal.npy
    | val
  | scans_test_sample (preprocessed ScanNet test split)

Run training

Modify the "dataset_path" variable with yours in the config yaml.

The network is trained with a two-stage strategy. The whole training process takes ~6 days with 4 Nvidia V100 GPUs.

python train.py ./configs/scannet_stage1.yaml
python train.py ./configs/scannet_stage2.yaml

Citation

If you find our work useful in your research, please consider citing:

@InProceedings{Zhao_2021_ICCV,
    author    = {Zhao, Wang and Liu, Shaohui and Wei, Yi and Guo, Hengkai and Liu, Yong-Jin},
    title     = {A Confidence-Based Iterative Solver of Depths and Surface Normals for Deep Multi-View Stereo},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {6168-6177}
}

Acknowledgement

This project heavily relies codes from NAS and we thank the authors for releasing their code.

We also thank Xiaoxiao Long for kindly helping with ScanNet evaluations.