README.md

---
license: apache-2.0
---

# Preparing ISO

## Datasets

We provide the OccScanNet dataset files [here](https://huggingface.co/datasets/hongxiaoy/OccScanNet/tree/main), but you should agree the term of use of *ScanNet*, *CompleteScanNet* dataset.

For **a simplified way** to  prepare the dataset, you **just** download the `preprocessed_data` to `ISO/data/occscannet` as `gathered_data` and download the `posed_images` to `ISO/data/scannet`.

However, the *complete dataset generating process* is provided as followed:

### OccScanNet

1. Clone the official MMDetection3D repository.

```bash
git clone https://github.com/open-mmlab/mmdetection3d.git ISO_mm
```

2. Swith to `v1.3.0` version.

```bash
cd ISO_mm
git checkout v1.3.0
```

3. Download the *ScanNet* dataset following [instructions](https://github.com/open-mmlab/mmdetection3d/tree/v1.3.0/data/scannet) and place `scans` directory as `ISO_mm/data/scannet/scans`.

> :bulb: Note
>
> Recommend you create a `posed_images` directory at data disk and link the `scans` directory and `posed_images` directory to `data/scannet`, then run the following command.

4. In this directory, extract RGB image with poses by running 

```bash
python extract_posed_images.py --max-images-per-scene 100
```

> :bulb: Note
>
> Add `--max-images-per-scene -1` to disable limiting number of images per scene. ScanNet scenes contain up to 5000+ frames per each. After extraction, all the .jpg images require 2 Tb disk space. The recommended 300 images per scene require less then 100 Gb. For example multi-view 3d detector ImVoxelNet samples 50 and 100 images per training and test scene.

Then obtained the following directory structure. 

```
scannet
├── meta_data
├── posed_images
│   ├── scenexxxx_xx
│   │   ├── xxxxxx.txt
│   │   ├── xxxxxx.jpg
│   │   ├── intrinsic.txt
├── scans
├── batch_load_scannet_data.py
├── extract_posed_images.py
├── load_scannet_data.py
├── README.md
├── scannet_utils.py
```

5. Download original *CompleteScanNet*

The ground truth labels we used are from [SCFusion](https://github.com/ShunChengWu/SCFusion#generate-gt). Ground truth is available at [here](https://github.com/ShunChengWu/SCFusion#generate-gt).

The ground truth label should be placed as `ISO_mm/data/completescannet/gt`.

6. Reformulate *CompleteScanNet*

```bash
python preprocess_gt.py
```

The resulted directory is `ISO_mm/data/completescannet/preprocessed`.

Now, we obtained the following directory structure.

```
completescannet
├── gt
│   ├── scenexxxx_xx.ply
├── preprocessed
│   ├── scenexxxx_xx.npy
├── preprocess_gt.py
├── visualization.py
```

7. Create the *OccScanNet*

First, you should create a directories with name `preprocessed_voxels` and `gathered_data` in data disk and link them to the `ISO_mm/data/occscannet`.

```bash
python generate_gt.py
```

Now, we obtained the following directory structure.

```
occscannet
├── preprocessed_voxels
├── gathered_data
├── generate_gt.py
├── not_aligns.txt
├── wrong_scenes.txt
├── bad_scenes.txt
├── used_scannames.txt
```

### OccScanNet-mini

The scenes we used in OccScanNet-mini is reflected in the config file.