[18/10/2024] Due to NTU RGB+D License, 320 videos from NTU RGB+D need to be downloaded manually. Please visit ROSE Lab to access the data. We also provide a list of the 320 videos used in MVBench for your reference.

We introduce a novel static-to-dynamic method for defining temporal-related tasks. By converting static tasks into dynamic ones, we facilitate systematic generation of video tasks necessitating a wide range of temporal abilities, from perception to cognition. Guided by task definitions, we then automatically transform public video annotations into multiple-choice QA for task evaluation. This unique paradigm enables efficient creation of MVBench with minimal manual intervention while ensuring evaluation fairness through ground-truth video annotations and avoiding biased LLM scoring. The 20 temporal task examples are as follows.

Evaluation

An evaluation example is provided in mvbench.ipynb. Please follow the pipeline to prepare the evaluation code for various MLLMs.

Preprocess: We preserve the raw video (high resolution, long duration, etc.) along with corresponding annotations (start, end, subtitles, etc.) for future exploration; hence, the decoding of some raw videos like Perception Test may be slow.
Prompt: We explore effective system prompts to encourage better temporal reasoning in MLLM, as well as efficient answer prompts for option extraction.