NL-Refer: Video Object Understanding with Natural Language Instructions

NL-Refer is a video object-level instruction dataset built on top of VideoRefer-700K. It replaces visual prompts with natural language descriptions to indicate target objects for fine-grained understanding. The dataset was created by BBBBCHAN and was last updated on Hugging Face in May 2026.

Use Cases

• Training models for video object grounding based on natural language referring expressions.
• Benchmarking fine-grained video object understanding systems that process language instructions.
• Researching vision-language alignment for video content as described in the dataset's overview.

Strengths

• Built on the VideoRefer-700K foundation, suggesting a substantial underlying data scale.
• Focuses on natural language instructions, a key feature for human-AI interaction.

Limitations

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

Provenance

Source

BBBBCHAN on Hugging Face.

Collection Method

Built as a derivative of the VideoRefer-700K dataset.

Freshness

Last updated 2026-05-20 08:54:09; freshness should be verified.