Skip to content

Latest commit

 

History

History
105 lines (83 loc) · 3.69 KB

README.md

File metadata and controls

105 lines (83 loc) · 3.69 KB
Raw

2D & 3D Data Preprocessing

Overview

This document provides instructions for pre-processing both 3D and 2D data for different datasets, including

  • ScanNet
  • Matterport3D
  • nuScenes
  • Replica

One can also use the provided code as a reference for pre-processing your customized dataset.

Prerequisites

Environment

Before you begin, simply activate the openscene conda environment.

Alternatively, make sure the following package installed:

  • torch
  • numpy
  • plyfile
  • opencv-python
  • imageio
  • pandas
  • tqdm

Download the original dataset

Run the code

For preprocessing 3D point clouds with GT labels (except for Replica, no semantic labels), one can simply run:

python preprocess_3d_{$DATASET_NAME}.py

For preprocessing 2D RGB-D images, one can also simply run:

python preprocess_2d_{$DATASET_NAME}.py

Note: In the code, you might need to modify the following:

  • in_path: path to the original downloaded dataset
  • out_dir: output directory to save your processed data
  • scene_list: path to the list containing all scenes
  • split: choose from train/val/test data to process

The only exception is preprocess_2d_scannet.py. The usage can be found in the code.

Folder structure

Once running the pre-processing code above, you should have a data structure like below. Here we choose the processed ScanNet as an example:

data/
│
├── scannet_2d
│   │
│   ├── scene0000_00
│   │   ├── color
│   │   ├── depth
│   │   ├── pose
│   │   └── intrinsic (only for Matterport3D)
│   │
│   ├── scene0000_01
│   │   ├── color
│   │   ├── depth
│   │   ├── pose
│   │   └── intrinsic
│   │
│   └── ...
|   |
|   └── intrinsics.txt (fixed intrinsic parameters for all scenes)
│
└── scannet_3d
    │
    ├── train
    │   ├── scene0000_00.pth
    │   ├── scene0000_01.pth
    │   ├── ...
    │   ├── scene0706_00.pth
    │
    └── val
        ├── scene0011_00.pth
        └── ...

Customized dataset: Make sure that you have the same structure after preprocessing.

Suggestions & Troubleshooting

If you encounter any issues, refer to the following troubleshooting tips:

  • We downscale 2D images for all dataset, so make sure you set img_dim and original_img_dim correctly.
  • To reproduce the results shown in Table 3 and 5 in the paper, you need to change num_classes in preprocess_3d_matterport_K_num_classes.py, which corresponds to the top K number of object classes in NYU Class annotation.
  • Matterport3D has per-image intrinsic parameters, so we have a separate folder for it. All other datasets have the global intrinsic parameters for all images and all scenes.
  • No train/val/test split for Replica.
  • Under dataset, we provide for ScanNet and matterport the scene_list (the list for all scene names) and official .tsv files for label mapping.

For additional help, please refer to the code documentation or contact the author.