Skip to content

Latest commit

 

History

History

u_mvs_casmvsnet

README.md

U-MVS-CASMVSNET

About

CascadeMVSNet is utilized as the backbone. The pretrained weight is provided for a fast evaluation. The training code will be uploaded later.

Note

  • It is suggested to use pytorch 1.2.0-1.4.0. The newest ones like pytorch 1.9.0 may fail to reproduce the same results. Please check the environment before running the evaluation code.

How to use?

Environment

  • pytorch 1.3.0
  • torchvision 0.4.2
  • cuda 10.1
  • apex
    • Note: apex is an optional library to accelerate the training, though it may degrade the performance a little.
    • Commands for installing apex: git clone https://github.com/NVIDIA/apex; cd apex; python setup.py install;
    • For more information, please to NVIDIA's official repository: https://github.com/NVIDIA/apex.

The conda environments are packed in requirements.txt.

Evaluating pretrained model on Tanks&Temples benchmark

  • Download preprocessed dataset: Tanks&Temples.
  • After decompressing the files, the dataset is organized as follows:
root_directory
 |--advanced
      |--Auditorium (scene name 1)
      |--Ballroom (scene name 2)
      |-- ...
 |--intermediate
      |--Family (scene name 1)
      |--Francis (scene name 2)
      |-- ...
  • Modify the DATAPATH variable in scripts/test_tanks_pretrained.sh according to the exact path of the Tanks&Temples dataset in your machine.
  • Run bash scripts/test_tanks_pretrained.sh YOUR_DATA_SPLIT YOUR_GPU_ID to evaluate the pretrained model.
    • YOUR_DATA_SPLIT means which partition of Tanks&Temples dataset is used, for example: advanced, or intermediate.
    • YOUR_GPU_ID represents the id of the used GPU. In default, 0 can be used.
    • For example, bash scripts/test_tanks_pretrained.sh intermediate 0, bash scripts/test_tanks_pretrained.sh advanced 0.
    • After running this command, the generated 3D point clouds will be saved in ./outputs_tanks directory.
  • To submit the results to the official website of Tanks&Temples, please follow their official instructions to organize the files as follows:
upload_directory
 |--Auditorium.log (camera parameters in scene 1)
 |--Auditorium.ply (generated ply file from scene 1)
 |--Ballroom.log (camera parameters in scene 2)
 |--Ballroom.ply (generated ply file from scene 2)
 |-- ...
 |-- t2_submission_credentials.txt (please obtain this credential file from official website)
 |-- upload_t2_results.py (please download this script from official website)
  • Note:
    • The ply files of all scenes can be found in ./outputs_tanks.
    • The log files can be found in each scene directory from provided data in Tanks&Temples.
    • The t2_submission_credentials.txt and upload_t2_results.py should be obtained from official website： Tanks&Temples.
    • To submite the results, you can type the command: python upload_t2_results.py --group (intermediate/advanced/both). Select each one of intermediate, advanced, both if you want to upload the results of intermediate dataset or advanced dataset or both.
    • After submitting the results, you can click the evaluation button on the official site of Tanks&Temples and wait for several hours. The evaluation results will be sent to your registered e-mail box once the evaluation is finished.

Evaluating pretrained model on DTU benchmark

  • Download preprocessed dataset: DTU and unzip it as the $TESTPATH folder which is organized as follows:
unzipped_data_directory ($TESTPATH)
 |--dtu
     |-- scan1
     |-- scan4
     |-- ...
  • Modify the $TESTPATH in test_pretrained.sh according to the exact path in your machine.
  • Prepare the fusibile following the instructions in the fusion section of u_mvs_mvsnet.
  • Run the script: bash scripts/test_pretrained.sh high.
    • Hyperparameters: num_consistent, prob_threshold, disp_threshold can be modified by your own.
  • The results will be saved in ./outputs directory.
  • Run bash scripts/arange.sh.
  • Evaluate the perfomance on DTU benchmark following the instructions in the Benchmark results on DTU section of u_mvs_mvsnet.

Note

  • It is suggested to use pytorch 1.2.0-1.4.0. The newest ones like pytorch 1.9.0 may fail to reproduce the same results.