Breaking01 / AI
Study raises video AI object-counting accuracy to 65.0%
The paper reports improved counting in synthetic videos alongside limits in real-video evaluation.

Dongwu Qian reported object-counting accuracy of 65.0% in synthetic videos in a paper submitted to arXiv on 21 Oct. The study examines how video AI counts repeated appearances of the same object.
The paper reports an improvement from 46.7% accuracy. It uses training that identifies the intervals when an object appears. It also uses an auxiliary channel that passes individual events as symbols.
Qian reports that analysis of the model's internal representations identifies object presence within individual intervals with 93% accuracy. Decoding the total count, however, is confounded by video duration. The diagnosis is that counting information exists internally but does not reach the final answer.
The paper reports an evaluation of 18 real videos in REALCOUNT. The best accuracy in that evaluation is 55.6%. Transferring the symbol-based approach to real videos without additional training failed.
Qian describes an earlier explanation that models fail to understand that objects still exist after leaving the screen. The study finds that explanation incomplete. It identifies the gap between internal information and final answers as a target for improvement.
The paper has not yet published updated results from the REALCOUNT adaptation experiments being retrained with corrected prompts.