ClevrTex: A Texture-Rich Benchmark for Unsupervised Multi-Object Segmentation

ClevrTex is a synthetic benchmark of 50,000 training and 10,000 test images designed to challenge unsupervised multi-object segmentation models. Created by Laurynas Karazija at the University of Oxford using physically based rendering, it features scenes with 3-10 objects rendered with diverse textures and materials. The dataset aims to expose the limitations of current state-of-the-art models, which perform well on simpler data but fail on this textured, complex imagery.

Use Cases

• Benchmarking unsupervised segmentation algorithms based on the dataset's textured scenes.
• Analyzing model robustness to visual confounders based on the controlled scene complexity variants.
• Training models for texture-invariant object recognition based on the 60 cataloged materials.
• Studying the gap between synthetic and real-world visual complexity based on the photorealistic rendering.

Strengths

• Contains 50,000 training examples and a separate 10,000-image test set.
• Scenes are procedurally generated with 3-10 objects using a catalog of 60 materials.
• Includes controlled variants to probe specific model shortcomings.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for individual dataset variants is unknown, which may limit suitability assessment.
• Data is fully synthetic, which may limit direct applicability to real-world imagery.

Provenance

Source

University of Oxford

Collection Method

Synthetic scenes created using physically based rendering techniques.