data_preparation

Object-NeRF: Data Preparation

ToyDesk Dataset

You can download the data through this link.

Please note that the transforms_xxx.json comes with transformation matrix (from camera coordinate to world coordinate) in SLAM / OpenCV format (xyz -> right down forward), thus we need to change to NDC format (xyz -> right up back) when applying to standard NeRF training codes.

ScanNet Dataset

You can download the pre-processed files through this link.

Step 1: Generate ScanNet training frames

First, we generate training frames (RGB-D image, instance segmentation) from ScanNet files to NeRF-like format. For example, to generate scene00xx_00, you can run script like that:

python data_preparation/scannet_sens_reader/convert_to_nerf_style_data.py \
    --input data/scannet/scans/scene00xx_00/ \
    --output data/scannet/processed_scannet_00xx_00 \
    --instance_filt_dir data/scannet/scans/scene00xx_00/instance-filt

Please note that the generated transforms_xxx.json comes with transformation matrix (from camera coordinate to world coordinate) in SLAM / OpenCV format (xyz -> right down forward), thus we need to change to NDC format (xyz -> right up back) in our dataloader.

Step 2: Generate ScanNet scene point clouds

When enabling voxel based representation, we use script from NPCR to generate point clouds via depth lifting.

Step 3: Make sure the config file matches data path

We assume the config file as config/scannet_base_0113.yml. - root_dir should be matched to the generated data folder in Step 1. - pcd_path should be matched to the generated scene point cloud in Step 2. - scans_dir is pointing to the ScanNet's official unzipped scans. Note that the bounding box and scene transformation has been stored in scannet_train_detection_data, which can also be generated following VoteNet docs.

For convenience, you can directly download the preprocessed scannet_train_detection_data (Step 3) and scene point cloud (Step 2) from this link.