视觉图灵：从人机对抗看计算机视觉下一步发展

黄凯奇1,2，赵  鑫1，李乔哲1，胡世宇1

(1. 中国科学院自动化研究所智能系统与工程研究中心，北京 100190；

2.中国科学院脑科学与智能技术卓越创新中心，上海 200031)

DOI：10.11996/JG.j.2095-302X.2021030001

内 容 简 介

摘    要：计算机视觉一直是人工智能研究的热点方向，经过近60年的发展，已经在算法、技术和应用等方面取得了巨大的进步。近十年来，以大数据、大算力为基础的深度学习进一步推动计算机视觉走向大模型时代，但其算法适应能力仍然和人类存在较大差距。本文从视觉任务评估评测（评测数据集、评测指标、评估方式）出发，对计算机视觉的发展进行了总结，对现存的依赖大数据学习的计算机视觉发展问题进行了梳理和分析，从人机对抗智能评测提出了计算机视觉下一步发展方向：视觉图灵。最后对视觉图灵发展方向进行了思考和讨论，探讨了未来研究可能的方向。

关键词：计算机视觉；视觉图灵；评估评测；图灵测试；数据集

VisualTuring: the next development of computer vision in the view of human-computer gaming

HUANGKai-qi^1,2,  ZHAO Xin¹, LI Qiao-zhe¹,  HU Shi-yu¹

Abstract: Computer vision has gainedwide attention in the research of artificial intelligence. After nearly 60years of its development, great achievement has been made in aspect ofalgorithms, technologies, and applications. Over the past decade, deeplearning, which is on the basis of big data and huge computation power, has furtherushered computer vision in an era of large model. However, there remains a hugegap between algorithm adaptability and human beings. From the perspective ofvisual task evaluation (in terms of datasets, metrics, and methods), this papersummarized the development history of computer vision. In addition, asystematic analysis was conducted on the existing problems and obstacles forthe development of computer vision heavily dependent on big data learning.Based on the analysis, this paper argued that the visual Turing test could bethe next research direction of computer vision. Finally, the development of thevisual Turing test and its potential research were discussed.

Keywords: computer vision;visual Turing; evaluation of visual tasks; Turing test; datasets

图1  从视觉任务评估评测看计算机视觉发展

Fig. 1  Thedevelopment of computer vision from the perspective of visual task evaluation

更多链接：https://kns.cnki.net/kcms/detail/10.1034.T.20210507.1635.004.html