Dataset · OpenGVLab
MVBench benchmark for video understanding by multimodal AIs
OpenGVLab/MVBench
A set of 4,000 video questions across 20 tasks to test time and action perception by multimodal models.
What's inside
The dataset from the OpenGVLab laboratory includes 4,000 English multiple-choice questions covering 20 types of tasks on understanding the dynamics and temporal sequence of events in video. The authors formed the set by converting static tasks into dynamic ones based on ready-made video markup. Some videos (320 videos from the NTU RGB+D set) must be downloaded separately manually due to licensing restrictions.
What you can do with it
- Testing multimodal language models (MLLMs) on action understanding over time
- Evaluating video-VQA quality and classification of complex dynamic scenes
- Comparing video encoder architectures on tasks requiring analysis of the entire duration of the clip
- Tasks
- Visual question answering, Video classification
- Languages
- English
- Size
- 1K–10K rows
- Downloads this month
- 40.1K
- Likes
- 48
Models trained on it
How to load it
from datasets import load_dataset
ds = load_dataset("OpenGVLab/MVBench")