Multi-Object Datasets: Image Scenes with Segmentation and Generative Factors

Google DeepMind released this collection of multi-object image datasets featuring ground-truth segmentation masks and generative factors for representation learning. The repository was last updated in March 2026 and provides standardized benchmarks for object-centric machine learning.

Use Cases

• Evaluating disentangled representation learning using ground-truth generative factors
• Training instance segmentation models with pixel-level segmentation masks
• Benchmarking object-centric learning algorithms on multi-object scenes

Strengths

• Includes ground-truth generative factors for latent variable analysis
• Provides pixel-level segmentation masks for multi-object scenes
• Authored by Google DeepMind for representation learning research

Limitations

• Likely restricted to synthetic or controlled environments
• Lack of documentation on specific row counts or file formats

Provenance

Source

Google DeepMind

Collection Method

synthetic

Freshness

Last updated March 2026.