FunTHOR: Synthetic 3D Scenes with Functional Relation Annotations

FunTHOR is a synthetic dataset for functional 3D scene understanding, built on the AI2-THOR simulator. It provides part-level ground-truth geometry and dense, rule-based functional-relation annotations for 12 indoor scenes, together with posed RGB-D sequences for each scene. The dataset was introduced as a benchmark by leggedrobotics.

Use Cases

• Training models for functional relation prediction based on rule-based annotations like 'knife slices apple'.
• Benchmarking 3D scene understanding algorithms based on part-level ground-truth geometry.
• Developing robotic manipulation policies using posed RGB-D sequences from indoor scenes.

Strengths

• Provides dense, rule-based functional-relation annotations for objects in scenes.
• Includes part-level ground-truth geometry for 12 indoor scenes.
• Contains posed RGB-D sequences for each scene, offering multimodal data.

Limitations

• Dataset size, row count, and specific file formats are unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• Data is synthetic, generated from the AI2-THOR simulator, which may not fully reflect real-world complexity.

Provenance

Source

huggingface dataset by leggedrobotics, built on the AI2-THOR simulator.

Collection Method

Synthetically generated within a simulation environment.

Freshness

Last updated 2026-06-14 17:16:47; freshness should be verified.