Skip to content

DATA_PREPARATION.md

Latest commit

 

History

History
127 lines (105 loc) · 4.81 KB

DATA_PREPARATION.md

File metadata and controls

127 lines (105 loc) · 4.81 KB

Data Preparation 🗂️

This document provides detailed instructions for preparing the datasets used in our project. We use two main datasets: ShapeNet-Multiview and ScanNet v2.

Table of Contents

ShapeNet-Multiview Dataset 🖼️ (For Object-level Pretraining)

ShapeNet-Multiview dataset is a collection of multi-view renderings of 3D models in the ShapeNet dataset.

Our ShapeNet-Multiview dataset is based on the work of ShapenetRender_more_variation. We configured different parameters specifically for compatibility with our framework. It provides high-quality multi-view renderings of 3D models with the following features:

  • 24 views per model
  • Resolution: 128 x 128
  • Modalities: RGB and point cloud
  • Camera parameters and metadata included

Dataset Download

Option 1: Download from Baidu Cloud

Link: https://pan.baidu.com/s/1XIuxMYMhXeIhd9Bf8XZuoQ?pwd=ve54
Extraction Code: ve54

Option 2: Download from HuggingFace

https://huggingface.co/datasets/Yanran21/Shapenet_multiview

Dataset Structure

The dataset is split into multiple zip files, which need to be downloaded, merged, and extracted. The commands for merging and extracting are as follows:

zip -s 0 shapenet_dataset.zip --out shapenet_dataset_merged.zip
unzip shapenet_dataset_merged.zip

After downloading and extracting, your dataset should follow this structure:

shapenet_dataset_merged/
├── image/
│   ├── 02691156/   
│   │   ├── 10155655850468db78d106ce0a280f87/
│   │   │   ├── easy/
│   │   │   │   ├── 00.png
│   │   │   │   ├── 00.txt
│   │   │   │   ├── 01.png
│   │   │   │   ├── 01.txt
│   │   │   │   ├── ...
│   │   │   │   ├── rendering_metadata.txt
│   │   │   ├── pts/
│   │   │   │   ├── 02691156-10155655850468db78d106ce0a280f87.npy
│   │   └── ...
│   └── 02747177/
│       ├── ...

where rendering_metadata.txt contains the camera parameters and metadata for each view. pts contains the point cloud data. image contains the rendered images.

For using the dataset, you need to set the data.dataset_root in the configs/dataset/shapenet.yaml file.

ScanNet v2 🏠 (For Scene-level Pretraining)

ScanNet v2 is a large-scale indoor scene dataset with RGB-D scans and 3D reconstructions.

Download and Preparation

We utilize both point clouds and RGB images from ScanNet v2. For point cloud, we employ the processed ScanNet data provided by Pointcept. For RGB images, we use the original ScanNet v2 dataset directly.

ScanNet Dataset Structure

After processing, your ScanNet point cloud dataset should look like:

scannet_pts_dataset_root/
├── train/
│   ├── scene0000_00/
│   │   ├── color.npy
│   │   ├── coord.npy
│   │   ├── normal.npy
│   │   ├── instance.npy
│   │   ├── segment20.npy
│   │   └── segment200.npy
│   └── ...
├── val/
│   ├── scene0000_00/
│   │   ├── color.npy
│   │   └── ...
│   └── ...
├── test/
│   ├── scene0000_00/
│   │   ├── color.npy
│   │   └── ...
│   └── ...

And your ScanNet rgb dataset should look like:

├── 2D
│   ├── color/
│   │   ├── scene0000_00/
│   │   │   ├── 000000.jpg
│   │   │   ├── 000001.jpg
│   │   │   ├── ...
│   │   └── ...
│   ├── depth/
│   │   ├── scene0000_00/
│   │   │   ├── 000000.png
│   │   │   ├── 000001.png
│   │   │   ├── ...
│   │   └── ...
│   ├── pose/
│   │   ├── scene0000_00/
│   │   │   ├── 000000.txt
│   │   │   ├── 000001.txt
│   │   │   ├── ...
│   │   └── ...

For using the dataset, you need to set the data.pts_dataset_root and data.rgb_dataset_root in the configs/dataset/scannet.yaml file.