Weighted Random Forest using Gaze Distributions Measured from Observers for Gender Classification

We weight random forest training using gaze distributions measured from observers performing a gender classification task on pedestrian images. The method aims to use features from regions informative for gender classification while reducing reliance on background bias.

A system trained to classify gender from person images may mistakenly rely on the background when certain backgrounds appear more often in one category. It can then perform poorly on images with different backgrounds. We investigated whether the places people look while judging gender can help address this problem.

We combine observers' gaze locations into a map and give greater importance to regions that attract more attention. The classifier is a random forest, which combines many smaller decision trees. The goal is to make these trees use information from the regions people look at.

We compared three ways of adding gaze-based weights: when choosing image features, when using their numerical values, and when evaluating decision rules. On a public dataset, weighting information gain—the measure of how much clearer a classification becomes after a decision split—was effective. It improved accuracy over the existing random forest methods used for comparison, with statistical analysis supporting the difference.

• 山口 紗也加, 西山 正志, 岩井 儀雄,
  視線位置分布を用いた重み付きランダムフォレストによる性別認識,
  電子情報通信学会論文誌 D, Vol.J102-D, No.8, pp. 495 - 505, August 2019.
• Sayaka Yamaguchi, Masashi Nishiyama, Yoshio Iwai,
  Weighted Random Forest using Gaze Distributions Measured from Observers for Gender Classification, Proceedings of 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP), Vol. 5: VISAPP, pp. 273 - 280, February 2019.
• 山口 紗也加, 吉村 宏紀, 西山 正志, 岩井 儀雄,
  視線位置分布を用いたランダムフォレスト改良による人物性別の認識,
  画像センシングシンポジウム (SSII), IS3-28, pp. 1 - 5, June 2018. [優秀学術賞]