Skip to content

Latest commit

 

History

History

sem_seg

README.md

Semantic segmentation of indoor scenes

Dataset

  1. Download prepared HDF5 data for training:
sh +x download_data.sh
  1. Download 3D indoor parsing dataset (S3DIS Dataset) for testing and visualization. "Stanford3dDataset_v1.2_Aligned_Version.zip" of the dataset is used. Unzip the downloaded file into "deep_gcns/data" and merge with the folder Stanford3dDataset_v1.2_Aligned_Version which already contains the patches S3DIS_PATCH.diff and DS_STORE_PATCH.diff then run,
cd ../data/Stanford3dDataset_v1.2_Aligned_Version
git apply S3DIS_PATCH.diff
find . -name ".DS_Store" -delete

Thanks to (loicland) for the fix.

Next run,

python collect_indoor3d_data.py

to generate "deep_gcns/data/stanford_indoor3d"

To generated your own HDF5 files, run,

python gen_indoor3d_h5.py

Train

We use 6-fold training, such that 6 models are trained leaving 1 of 6 areas as the testing area for each model. We keep using 2 GPUs for distributed training. To train 6 models sequentially, run

sh +x train_job.sh

If you want to train model with other gcn layers (for example mrgcn), run

MODEL="ResMRGCN-28"
python train.py --log_dir $MODEL/log5 --test_area 5 --gcn 'mrgcn'

Evaluation

Make sure you have successfully setup the dataset for testing (Dataset, Step 2). You should have .npy files with the dataset in the folder "deep_gcns/data/stanford_indoor3d" and auxilary files pointing to the data in the folder "meta".

If you want to evaluate across all 6 areas (6-fold cross-validation):

  1. To generate predicted results for all 6 areas, run
sh +x test_job.sh
  1. Obtain overall quantitative evaluation results
python eval_iou_accuracy.py

If you want to evaluate on one area (for example area5):

  1. To generate predicted results for this area (area5), run
python batch_inference.py --model_path $MODEL/log5/epoch_$EPNUM.ckpt --dump_dir $MODEL/log5/dump --output_filelist $MODEL/log5/output_filelist.txt --room_data_filelist meta/area5_data_label.txt
  1. To obtain the quantitative evaluation results on this area, run
python eval_iou_accuracy.py --folder $MODEL --area 5

Pretrained Models

Several pretrained models (ResGCN-56, ResGCN-28W, ResGCN-28, DenseGCN-28, PlainGCN-28) can be downloaded:

  1. To install gdown which is required to download the pretrained models, run
pip install gdown
  1. To download the pretrained models and put them into the folder "pretrained", run
sh download_pretrained_models.sh

If you want to evaluate one of these models:

  1. Select the corresponding model in the file "test_pretrained.sh"

  2. To generate the predicted results and evaluate on area 5, run

sh +x test_pretrained.sh

Visualization

In order to compare several pretrained models visually you can follow these steps:

  1. To generate the predicted results including files for visualization, run
sh +x visualize_pretrained.sh
  1. Open the visualization notebook and follow the steps in the notebook.
jupyter notebook visualization.ipynb