Computer Vision Falls Short of Humans | Generated by AI

Home PDF

The blog post by Yin Wang argues that computer vision systems, even advanced ones in 2019, were far from achieving human-level performance. It claims that these systems struggle with basic object recognition (e.g., identifying a car) and are easily fooled by adversarial examples or minor image alterations, unlike humans who recognize objects effortlessly. Wang suggests that the field overstates its progress and that true human-level computer vision remains elusive due to fundamental limitations in how these systems process and understand images.

Is it true?

As of the post’s publication in October 2019, Wang’s argument had merit based on the state of computer vision at the time:

However, the post’s tone is absolute, claiming “there is no human-level computer vision.” This overlooks progress in specific tasks. For instance:

Critical Assessment

Wang’s core claim—that computer vision in 2019 was not human-level—is largely true. Models lacked the generalization, robustness, and intuitive understanding of human vision. However, his dismissal of progress may be overly pessimistic, as significant strides have been made since. Even in 2025, while computer vision excels in specific domains, it still falls short of human-level perception in open-world scenarios due to issues like:

The post remains relevant as a critique of overhyping AI capabilities but doesn’t account for the rapid progress in the field. No single source confirms “human-level” vision has been achieved by 2025, but the gap has narrowed significantly.


Back 2025.06.28 Donate