[2025/07/22]
Since errors in a few clue annotations when converting frame indexes to timestamps, there were errors in the previous benchmark leaderboard, we have reevaluated all models and have updated the new leaderboard.
Despite progress in video understanding, current MLLMs struggle with counting tasks. Existing benchmarks are limited by short videos, close-set queries, lack of clue annotations, and weak… See the full description on the dataset page:
https://huggingface.co/datasets/CG-Bench/CG-AV-Counting.
