Synthetic Lego Brick Images for Instance Segmentation

1,000 synthetic images of Lego bricks were procedurally generated using a Houdini-based pipeline. The dataset, created by author nathankimnguyen412, was last updated in March 2026. Annotations for instance segmentation and object detection were extracted automatically via Cryptomatte AOVs without manual labeling.

Use Cases

• Train instance segmentation models using automatically generated pixel-level annotations from Cryptomatte AOVs.
• Benchmark object detection algorithms on the single 'Lego brick' class across 1,000 rendered images.
• Develop domain adaptation techniques by testing models trained on synthetic data against real-world images.
• Evaluate the performance of rendering pipelines like Houdini Karma XPU for generating training data.

Strengths

• 1,000 procedurally generated images provide a controlled, scalable data source.
• Annotations are automatically extracted via Cryptomatte AOVs, eliminating manual labeling effort and potential human error.

Limitations

• The dataset contains only 1,000 images, which is a relatively small sample size for training deep learning models.
• All data is synthetic, which may not capture the full complexity and noise of real-world images, leading to a domain gap.
• Only a single class ('Lego brick') is represented, limiting the scope of multi-class recognition tasks.

Provenance

Source

nathankimnguyen412 on Hugging Face and GitHub.

Collection Method

Procedurally generated using a Houdini-based data generation pipeline and rendered with Karma XPU in Houdini Solaris.

Freshness

Last updated March 2026.