Blender Synthetics

Create your own synthetic dataset for object detection.

Description

This repo creates random scenes on blender, imports custom 3D models with randomised configurations, and renders numerous images with annotations to train an object detection model.

Getting Started

Install blender (Tested on 3.2 and 3.3)

• Follow installation instructions here
• Add blender to PATH
  • echo 'export PATH=/path/to/blender/directory:$PATH' >> ~/.bashrc
• Ensure your GPU is supported on blender. Refer here

Clone repo

git clone https://github.com/wish2023/blender-synthetics.git

Install packages

sh install_requirements.sh

Generate synthetics

• Update config/models.yaml and config/render_parameters.yaml as required. Refer to models and render parameters for details.
• Generate images
  • blender -b -P src/render_blender.py
• Generate annotations
  • python3 src/create_labels.py

Models

Currently supports fbx/obj/blend. Ensure your models only contain one object that has the same name as its filename.

Classes

Your targets of interest. Bounding boxes will be drawn around these objects.

Obstacles (Optional)

Other objects which will be present in the scene. These won't be annotated.

Scenes (Optional)

Textures that your scene may have. Explore possible textures from texture haven and store all texture subfolders in a main folder.

Parameters

Occlusion awareness

When not occlusion aware, bounding boxes will surround regions of the object that aren't visible by the camera.

Visibility threshold

The fraction of an object that must be visible by the camera for it to be considered visible to a human annotator.

Component visibility threshold

The fraction of an object components that must be visible by the camera for it to be considered visible to a human annotator.

Camera configurations

Sun

The sun's energy is the light intensity on the scene. The tilt is responsible for casting shadows and works similar to the camera's tilt.