- Download the ScanNet V2 dataset and go to its dir, which has the following files
scannetv2-labels.combined.tsv # label for train and val set
scans # scans for train and val set
scans_test # scans for test set
tasks- Preprocess the train and val files as follows
python -u scannet/preprocess_scannet.py \
--input PATH_TO_YOUR_SCANNETV2/scans \
--output SOME_PATH/scannet_fully_supervised_preprocessed \
--splits train # train includes both train & val- Copy the split files to your preprocessed dir
cp -r scannet/splits SOME_PATH/scannet_fully_supervised_preprocessed/- Generate the label ids, which are used to provide the
weak (sparse) labels - Two types of sparse labels:
- Absolute points: 20,50,100,200 points from Official ScanNet Data Efficient Benchmark
- Percentage points: 0.01%, 0.1% generated by our script (:fire: the script will be released later)
cp -r scannet/points SOME_PATH/scannet_fully_supervised_preprocessed/
- Download S3DIS data by filling this Google form. Download the Stanford3dDataset_v1.2.zip file and unzip it.
- Run preprocessing code for S3DIS as follows:
python prepare_dataset/stanford/preproceess_stanford.py --input ${PATH_TO_S3DIS_DIR} --output_root ${PROCESSED_S3DIS_DIR}- Copy the split files to your preprocessed dir
cp -r prepare_dataset/stanford/splits SOME_PATH/stanford_fully_supervised_preprocessed/- Generate the label ids, which are used to provide the
weak (sparse) labels - Only one type of sparse labels:
- Percentage points: 0.01%, 0.02%, 0.1%, 0.2%, 10%, generated by our script (:fire: the script will be released later)
cp -r prepare_dataset/stanford/points SOME_PATH/stanford_fully_supervised_preprocessed/🔥 Coming soon!