Stream3D-Bench is a benchmark with 10,000 question-answer samples for evaluating online 3D spatial understanding in vision-language models. It was introduced by author JonnyYu828 in 2026. The benchmark focuses on a streaming setting where models must process temporally ordered observations, unlike conventional offline 3D understanding tasks.
Use Cases
- Benchmarking model performance on online 3D spatial reasoning based on the streaming evaluation setting.
- Training models to process incremental geometry priors based on the description of the associated Stream3D-VLM model.
- Evaluating vision-language model robustness in temporally ordered scene observations based on the benchmark's core design.
Strengths
- Contains 10,000 question-answer samples for evaluation.
- Designed for a specific, modern task of online 3D spatial understanding in a streaming setting.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment.
- Data may reflect temporal or source bias inherent to the collection method.
Provenance
- Source
- huggingface
- Collection Method
- Introduced with the Stream3D-VLM research; likely contains synthetically generated or curated 3D scene data.
- Freshness
- Last updated 2026-06-02 14:39:10; freshness should be verified.