3D-FRONT Dataset

1. Download 3D-FRONT dataset from https://tianchi.aliyun.com/specials/promotion/alibaba-3d-scene-dataset

├──RAW_DATA_PATH
    ├── 3D-FRONT
    ├── 3D-FUTURE-model
    ├── categories.py
    ├── model_info.json

Note: 3D-FRONT is updated after the paper submission so the preprocessing is a little different from the original paper (marked as DFP below).

2. Extract the top 20 classes (Optional)

cd dataset_3dfront

• Generate assets/cat2id_all.pkl:

python gen_cat2id_all.py # assets/gen_cat2id_all.py

• Generate assets/cat2id_bedroom.pkl and assets/cat2id_living.pkl

type=bedroom; # or living
python json2layout_distribution.py --type $type

3. Generate the datasets for all room types

• Generate the split dataset:

type=bedroom; # or living
python json2layout.py --type $type --future_path RAW_DATA_PATH/3D-FUTURE-model --json_path RAW_DATA_PATH/3D-FRONT --model_info_path RAW_DATA_PATH/model_info.json

4. Filter and merge the datasets for each room type

For bedroom, we merge Bedroom.npy, MasterBedroom.npy, SecondBedroom.npy together. We filter out bedroom without bed. 4111 scenes in total. we use all for training (4000) and validation (111).

python mergeDataset_bedroom.py 

For livingroom, we merge LivingRoom.npy, LivingDiningRoom.npy together. 4684 scenes in total. We use all for training (4000) and validation (684, DFP). We filter out scenes with less than 6 objects (DFP 4 in the paper).

python mergeDataset_living.py 

Bedroom_train_val.npy and Livingroom_train_val.npy are generated in ./outputs .

5. Generate the final abs & rel dataset

type=bedroom; # or living
python gen_3dfront_dataset.py --type $type

The final dataset is stored in ./data/bedroom or ./data/living .

Final dataset:

├──data
    ├── bedroom
       ├── 0_abs.npy
       ├── 0_rel.pkl
       ├── ...
    ├── living
       ├── 0_abs.npy
       ├── 0_rel.pkl
       ├── ...
    ├── train_bedroom.txt
    ├── train_living.txt
    ├── val_bedroom.txt
    └── val_living.txt