MVC: Multi-View Captions for InternVid-10M-FLT

This dataset contains caption annotations for approximately 10 million clips from the InternVid-10M-FLT subset. It follows the multi-view supervision framework described in Advancing Video-Text Pretraining with Multi-View Captions.

The file data/internvid_10m_flt_all_captions.json is a JSON array with 10,636,927 records. Each record contains a source clip identifier or filename in video, along with the caption fields below. This repository contains annotations and clip identifiers; it does not contain the source video files. Access to the corresponding clips is governed by the terms of the upstream InternVid dataset.

Caption views

The paper organizes supervision into an original caption, concise summary captions, refined detailed captions, and semantic positive captions. The fields in this export are grouped accordingly.

Original caption

• original_internvid: The original caption associated with the clip in InternVid-10M-FLT (denoted O in the paper).

Summary and activity captions

The paper's summary view provides a concise semantic anchor focused on the dominant activity. This export contains three short activity or summary variants:

• tarsier_activity: A concise activity-focused caption from the Tarsier2-Recap-7B pipeline.
• tarsier_summary: A concise summary from the Tarsier2-Recap-7B pipeline.
• qwen3vl_activity: A concise activity-focused caption from the Qwen3-VL-30B-Instruct pipeline.

These fields complement the original caption with concise descriptions of the clip. The paper denotes its two model-generated summary views by (S_1) and (S_2); this export retains the three short activity or summary fields listed above.

Refined detailed captions

Detailed captions provide broader visual context, including visible objects, actions, interactions, and scene details. The paper refines each detailed draft by checking it against the video, correcting unsupported content and removing irrelevant overlays or non-visual inferences.

• denoised_detailed_qwen: A Qwen3-VL-30B-Instruct detailed draft refined by Qwen3-VL-8B-Thinking.
• denoised_detailed_tarsier: A Tarsier2-Recap-7B detailed draft refined by Qwen3-VL-8B-Thinking.

These fields correspond to the paper's refined detailed captions, denoted (D^*). The unrefined detailed drafts are intermediate outputs and are not included in this file.

Semantic positive captions

• pos1: First semantic positive caption.
• pos2: Second semantic positive caption.

Qwen3-VL-8B-Thinking generates these captions using the video and one selected refined detailed caption. They are not intended as simple paraphrases: each describes a visually grounded aspect of the clip, such as an actor, action, interacting object, or outcome, while providing a complementary semantic focus. They form the semantic-positive supervision, denoted (P_1) and (P_2) in the paper.

Record fields

Field	Description
video	Source clip identifier or filename used to match the annotation to an InternVid clip.
original_internvid	Original InternVid-10M-FLT caption.
tarsier_activity, tarsier_summary, qwen3vl_activity	Concise activity or summary views.
denoised_detailed_qwen, denoised_detailed_tarsier	Refined detailed captions.
pos1, pos2	Semantic positive captions.

Intended use and limitations

The annotations are intended for video-text research, including pretraining and retrieval. The paper uses separate summary and detailed text views so the model can learn complementary caption granularities. Captions are model-generated; refinement improves visual grounding but may leave residual inaccuracies. Check the source video when correctness matters.

Use the underlying InternVid-10M-FLT data in accordance with its original access terms and license.

Citation

@article{thoker2026mvc,
  title   = {Advancing Video-Text Pretraining with Multi-View Captions},
  author  = {Thoker, Fida M. and Vandeghen, Renaud and Sanchez, Karen and Van Droogenbroeck, Marc and Ghanem, Bernard},
  journal = {arXiv preprint arXiv:2609.35090},
  year    = {2026}
}