Stream3D-1M: Over 1 Million 3D Question-Answer Pairs from RGB-D Video Streams

Stream3D-1M-Dataset is a large-scale online spatio-temporal 3D question-answering dataset for training vision-language models. It contains over 1 million QA pairs generated from RGB-D video streams to support real-time 3D spatial understanding, temporal memory, and interactive reasoning. The dataset was introduced by author JonnyYu828 and was last updated on Hugging Face in June 2026.

Use Cases

• Training models for online 3D spatial understanding based on RGB-D video streams.
• Developing temporal memory capabilities for vision-language models based on streaming video data.
• Enabling interactive reasoning in 3D environments based on the described QA pairs.
• Benchmarking incremental geometry priors for 3D vision-language tasks.

Strengths

• Contains over 1 million question-answer pairs, providing a large-scale resource.
• Designed specifically for online, real-time 3D understanding from streaming data.

Limitations

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

Provenance

Source

JonnyYu828 on Hugging Face

Collection Method

Generated from RGB-D video streams; specific generation method not detailed in provided description.

Freshness

Last updated 2026-06-02 14:39:59; freshness should be verified.