MATT-Bench: Fine-Grained Mistake Attribution in Egocentric Videos

Mistake Attribution (MATT) benchmarks from CVPR 2026 go beyond binary mistake detection to attribute semantic role violations and identify Points-of-No-Return. The dataset, created by researchers from the University of Michigan and Voxel51, provides large-scale benchmarks for fine-grained mistake understanding in first-person videos. It was last updated on April 17, 2026.

Use Cases

• Train models for fine-grained mistake attribution based on semantic role violations mentioned in the description.
• Benchmark models for identifying the Point-of-No-Return in procedural mistakes from video.
• Develop video understanding systems for detailed error analysis in first-person activities.

Strengths

• Large-scale benchmark for a novel computer vision task.
• Provides annotations for multiple aspects of mistakes, including semantic roles and irreversibility.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

University of Michigan, Voxel51

Freshness

Last updated 2026-04-17 07:57:35; freshness should be verified.