Custom dataset

Usage

1. Follow the instruction in SAM(https://github.com/facebookresearch/segment-anything) to download the checkpoint and install SAM with

pip install git+https://github.com/facebookresearch/segment-anything.git

We use the vit-h model to generate our data.

2. Run SAM to get the segmentation data:

python seg.py --input $DATASET_PATH/{train, val} --output $DATASET_PATH/{train, val}_seg --model-type vit_h --checkpoint $CKPT_PATH --resize-fact $RESIZE_FACT$

This will generate segmentation labels in the following structure:

{train, val}_seg
    - IMG_NAME
        - MASK_NAME
        ...
    ...

We recommend resizing the image to the size used for NeRF training before running SAM to make the data processing more efficient.

3. Get the segmentation hierarchy with:

python get_hierarchy.py --input $DATASET_PATH/{train, val}_seg --output $DATASET_PATH/{train, val}_seg_hierarchy