Render ShapeNet Dataset

This script will help you render ShapeNet and custom datasets that follow ShapeNet conventions.

Prerequisites

Python 3.7 or higher
Blender 2.9 or higher
OSX, Linux or Windows running WSL2

Setup

Download the ShapeNet V1 or V2 dataset following the official link
Make a new folder called shapenet in this directory, and unzip the downloaded file: mkdir shapenet && unzip SHAPENET_SYNSET_ID.zip -d shapenet
Download Blender following the official link

Installing Required Libraries

You will need the following libraries on Linux:

apt-get install -y libxi6 libgconf-2-4 libfontconfig1 libxrender1

Blender ships with its own distribution of Python, which you will need to add some libraries to:

cd BLENDER_PATH/2.90/python/bin
./python3.7m -m ensurepip
./python3.7m -m pip install numpy

Example

from render_dataset import BlenderDataset
import cv2
ds = BlenderDataset("shapenet/cars/dataset_list.json", resolution=2048, shapenet_version="1")
img, condinfo, mask = ds[28]
img = img.transpose(1, 2, 0).astype(np.uint8)
cv2.imwrite("img.png", img)

Data

The rendering script looks for datasets in the dataset_list.json file. You can modify this to add your own files and paths or point to your own JSON dataset list using the --dataset_list <filename> flag when invoking render_all.py

For ShapeNetCore.v1, you don't need to do any preprocesing. If you are using your own dataset, you should make sure that your models are sorted into directories with a "model.obj" in them, following the expected conventions of ShapeNetCore.v1.
For ShapeNetCore.v2, make sure to pass the --shapenet_version 2 flag to the render_all.py script -- this will destructively normalize your dataset folder to match the expected structure of ShapeNetCore.v1, while retaining the original .obj and .mtl file names

Rendering

Quick Start

Once you've modified dataset_list.json or added the ShapeNet data that reflects the source training data paths, you can render your data like this:

python render_all.py

Note: The code will save the output from blender to tmp.out, this is not necessary for training, and can be removed by rm -rf tmp.out

Additional Flags

You can customize the rendering script by adding flags.

Switch to Eevee for dramatically faster rendering speed

By default, the Blender renderer uses Cycles, which has a photorealistic look but is slow (>10s/frame). You can also use Eevee, which may be more game-like in look but renders much much faster (<.3s/frame), and may be suitable for extracting a high quality dataset on lower end machines in a reasonable amount of time.

python render_all.py --engine EEVEE

Render ShapeNet V2

For ShapeNetCore.v2 you will need to pass a flag to the render script to pre-process your data:

python render_all.py --shapenet_version 2

Log to Console

By default, the script will log to a tmp.out file (quiet mode), but you can override this:

python render_all.py --quiet_mode 0

Set Number of Views to Capture

The default for the rendering script is to capture 24 views per object. However, many NeRF pipeline recommend closer to 100 images. Especially if you are working with a limited but high quality dataset, you should consider increasing the total number of views 2-4x

python render_all.py --num_views 96

Override Arguments

By default, the rendering script will save outputs to "shapenet_rendered", read all datasets from dataset_list.json and use the default Blender installation in your system. However, you can override these arguments:

python render_all.py --save_folder PATH_TO_SAVE_IMAGE --dataset_list PATH_TO_DATASET_JSON --blender_root PATH_TO_BLENDER

Modifying the Render Scene

You can open the base scenes (located in the blender directory) and modify the lighting. There are no objects in the scene, so you will need to import a test object. Just be careful to remove any scene objects before you save.

Comparison Between Cycles and Eevee

Cycles is on the Left, Eevee is on the Right

To Render Eevee headlessly

!apt-get install python-opengl -y
!apt install xvfb -y
!pip install pyvirtualdisplay
!pip install piglet
python3 render_parallel.py --num_views 96 --engine EEVEE --headless

Attribution

This code is adopted from this GitHub repo, we thank the author for sharing the codes!
The tome in the rendering comparison images was borrowed with permission from the Loot Assets library.
Original by lalalune.

neverix/BlenderDataset