R3D-Bench: Quantitative 3D Spatial Reasoning from Egocentric RGB-D Video

Facebook provides 3,033 question-answer annotations over 57 Aria Digital Twin sequences for evaluating quantitative 3D spatial reasoning from natural egocentric RGB-D video. This is an evaluation-only benchmark with no training split; all data is designated for testing. The underlying Aria video data must be downloaded separately.

Use Cases

• Benchmarking 3D spatial reasoning models based on question-answer annotations
• Evaluating AI performance on quantitative tasks from egocentric RGB-D video
• Developing visual question answering systems for 3D environments based on the described annotations

Strengths

• 3,033 question-answer annotations provide a substantial evaluation set
• Covers 57 distinct Aria Digital Twin sequences for varied scenarios

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count for the primary annotation table is unknown, which may limit suitability assessment

Provenance

Source

facebook

Collection Method

Annotations created over Aria Digital Twin sequences.

Freshness

Last updated 2026-07-03 01:36:16; freshness should be verified