Home

Awesome

scannet_toolkit

Scripts for downloading a smaller random sample of the ScanNet dataset and generating depth maps from the PLY models of the scenes.

Important notes

Dependencies

Installation

Clone the repo recursively:

git clone --recurse-submodules https://github.com/jczarnowski/scannet_toolkit.git

Obtain the download-scannet.py by requesting access to the dataset and following the instructions from the authors' website. Place this script in the root directory of this repository.

Usage

We used this repo for sampling a subset of the ScanNet dataset and preprocess it in the following way:

  1. Create random samples for train and val with sample.sh
  2. Download both samples with download_sample.sh
  3. Preprocess all RGB images to the same size as the depth (640x480). We had to do this as the colour image sizes are inconsistent accross the dataset and our preprocessing code requires that
  4. Render depth using the supplied PLY scene models and ground truth poses with render.py. This gives us a nice smooth model to fill the gaps in original depth maps
  5. Merge the original depth maps with the rendered ones using merge_depths.py. Thanks to this step, we fill the gaps in the depth maps with the model and model gaps with raw data.

Sampling

First, generate a random sample of scenes using sample.sh. To generate a 40% sample of the ScanNet training set:

bash sample.sh 40 samples/scannetv2_train.txt > samples/my_40_sample.txt

Files samples/scannetv2_train.txt and samples/scannetv2_val.txt contain the scene names for the official ScanNet test/train split.

If you are sampling from the same file, you should also make sure that your generated train/test split is exclusive. This can be quickly done with e.g. grep:

grep -Fxf file1 file2

Downloading

Download your sample using download_sample.sh to a directory out_dir as follows:

bash download_sample.sh samples/my_40_sample.txt <out_dir>

The script currently only pulls and decodes the .sens files and the _vh_clean_2.ply models.

Preprocessing colour images

The preprocess.py script is a WIP and resizes the images to 640x480.

Rendering depth from models

The render.py script uses the trajectory and the scene PLY models to render completed depth maps and normals. Example usage:

python render.py --data_dir <out_dir>

Rendered depth and normals are saved alongside the images:

<sceneid>/frame-XXXXXX.rendered_depth.png
<sceneid>/frame-XXXXXX.rendered_normal.png

Depth is in milimeters and normals are converted to be in range [0,1] with: out_normal = (in_normal + 1) / 2.0

Merging rendered and raw depth

The merge_depths.py script reads in the rendered and the raw depth maps and combines them together into a single image and saves it alongside:

<sceneid>/frame-XXXXXX.rendered_depth.png
<sceneid>/frame-XXXXXX.rendered_normal.png
<sceneid>/frame-XXXXXX.merged_depth.png

Example usage:

python merge_depths.py --data_dir <out_dir>

Authors