Chaotic Object Scene for Blender Object Detection Datasets

A scene you can re-use for quickly generating annotated images for object detection tasks.

Primary purpose: Generate varied images on a custom object, with the goal of fine-tuning and object detector model (like YOLO).

The scene provides plenty of distractor objects, lighting variation, object variance, and chances for occlusion. Using images from this scene can quickly get you to a baseline object detector with popular models like YOLO.

The best way to use this scene is with the free Bl-Vision Add-On

For an article covering the motivation behind this add-on as well as some background:

Towards an Accessible Way of Creating Synthetic Images for Vision Models

The scene is based on the particle system Blender has built in. To use with your own model:

• Insert an object of your choosing
• Select the Target Emitter object and go to the Particle Settings
• Replace the rendered object (currently a 3D Benchy) with your object.

To generate labels automatically:

• Install the Bl-Vision Blender Add-On
• Then simply select the Target Emitter (which should be selected by default with this scene).
• Select your target root folder
• Ensure Bounding Box checkmark is checked.
• Hit Render.

The scene is inspired on a paper from NVIDIA, which can be found here:

[1] R. Singh et al., “Synthetica: Large Scale Synthetic Data for Robot Perception,” Oct. 28, 2024, arXiv: arXiv:2410.21153. doi: 10.48550/arXiv.2410.21153.

However, unlike the paper, it is meant for quickly fine-tuning object detection models (like YOLO).