3DMV jointly combines RGB color and geometric information to perform 3D semantic segmentation of RGB-D scans. This work is based on our ECCV'18 paper, 3DMV: Joint 3D-Multi-View Prediction for 3D Semantic Scene Segmentation.

Training is implemented with PyTorch. This code was developed under PyTorch 0.2 and recently upgraded to PyTorch 0.4.

• See python train.py --help for all train options. Example train call:

python train.py --gpu 0 --train_data_list [path to list of train files] --data_path_2d [path to 2d image data] --class_weight_file [path to txt file of train histogram] --num_nearest_images 5 --model2d_path [path to pretrained 2d model]

• Trained models: models.zip

• See python test.py --help for all test options. Example test call:

python test.py --gpu 0 --scene_list [path to list of test scenes] --model_path [path to trained model.pth] --data_path_2d [path to 2d image data] --data_path_3d [path to test scene data] --num_nearest_images 5 --model2d_orig_path [path to pretrained 2d model]

This data has been precomputed from the ScanNet (v2) dataset.

• Train data for ScanNet v2: 3dmv_scannet_v2_train.zip (6.2G)
  • 2D train images can be processed from the ScanNet dataset using the 2d data preparation script in prepare_data
  • Expected file structure for 2D data:

  scene0000_00/
  |--color/
     |--[framenum].jpg
         ⋮
  |--depth/
     |--[framenum].png   (16-bit pngs)
         ⋮
  |--pose/
     |--[framenum].txt   (4x4 rigid transform as txt file)
         ⋮
  |--label/    (if applicable)
     |--[framenum].png   (8-bit pngs)
         ⋮
  scene0000_01/
  ⋮
  

• Test scenes for ScanNet v2: 3dmv_scannet_v2_test_scenes.zip (110M)

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

@inproceedings{dai20183dmv,
 author = {Dai, Angela and Nie{\ss}ner, Matthias},
 booktitle = {Proceedings of the European Conference on Computer Vision ({ECCV})},
 title = {3DMV: Joint 3D-Multi-View Prediction for 3D Semantic Scene Segmentation},
 year = {2018}
}

If you have any questions, please email Angela Dai at [email protected].