CiNet Monthly Seminar
<\/p>\n\n\n\n
2026\u5e746\u670822\u65e5\uff08\u6708\uff09
15:00-16:00
CiNet\u68df\u5927\u4f1a\u8b70\u5ba4\u306b\u3066\u958b\u50ac<\/p>\n\n\n\n
\u6f14\u984c\uff1aAligning machine and human in visual motion perception<\/p>\n\n\n\n
\u4eac\u90fd\u5927\u5b66 \u5927\u5b66\u9662\u60c5\u5831\u5b66\u7814\u7a76\u79d1
\u6559\u6388
\u897f\u7530 \u771e\u4e5f<\/p>\n\n\n\n
\u62c5\u5f53 :\u00a0\u5c0f\u6797 \u52c7\u8f1d
Abstract:
One of my research goals is to advance our understanding of human perception of visual motion. Our main strategy was to use carefully controlled visual stimuli and psychophysical analyses. Building on these findings, recent work has focused on developing computational models that can comprehensively explain motion perception for a wide range of visual inputs, including natural movies.<\/p>\n\n\n\n
To this end, we constructed an artificial neural network model for motion flow estimation that combines motion-energy detection mechanisms analogous to those in V1 with a self-attention network inspired by motion integration processes in MT [1]. By training the model on a large set of video sequences, we succeeded in reproducing various perceptual characteristics of luminance-based motion (first-order motion). Furthermore, by training the model to estimate the motion of glossy objects, the network naturally acquired sensitivity to second-order motion cues such as texture-defined motion.<\/p>\n\n\n\n
In parallel, we estimated motion vectors perceived by human observers for several benchmark video datasets commonly used in optical-flow evaluation and training in machine vision studies [2, 3]. Comparisons with our model [1] showed that it could account for some aspects of human perceptual errors, but only to a limited extent. We hypothesize that this discrepancy arises because conventional motion-flow models estimate retinal motion vectors, whereas humans analyze the hierarchical structure of motion and perceive relative motion with respect to reference frames. Based on this idea, we have also begun investigating hierarchical vector decomposition as a framework for understanding higher-order motion processing in human vision.<\/p>\n\n\n\n
[1] Sun, Z., Chen, Y.-J., Yang, Y.-H., Li, Y., & Nishida, S. (2025). Machine learning modelling for multi-order human visual motion processing. Nature Machine Intelligence, 7, 1037\u20131052.<\/p>\n\n\n\n
[2] Yang, Y.-H., Fukiage, T., Sun, Z., & Nishida, S. (2023). Psychophysical measurement of perceived motion flow of naturalistic scenes. iScience, 26(12), 108307.<\/p>\n\n\n\n
[3] Yang, Y.-H., Sun, Z., Fukiage, T., & Nishida, S. (2025). HuPerFlow: A Comprehensive Benchmark for Human vs. Machine Motion Estimation Comparison. In Proc. IEEE\/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 22799\u201322808 (CVF, 2025).<\/p>\n","protected":false},"featured_media":0,"template":"","acf":[],"_links":{"self":[{"href":"http:\/\/cinetjp-static3.nict.go.jp\/japanese\/wp-json\/wp\/v2\/event\/6789"}],"collection":[{"href":"http:\/\/cinetjp-static3.nict.go.jp\/japanese\/wp-json\/wp\/v2\/event"}],"about":[{"href":"http:\/\/cinetjp-static3.nict.go.jp\/japanese\/wp-json\/wp\/v2\/types\/event"}],"version-history":[{"count":9,"href":"http:\/\/cinetjp-static3.nict.go.jp\/japanese\/wp-json\/wp\/v2\/event\/6789\/revisions"}],"predecessor-version":[{"id":6850,"href":"http:\/\/cinetjp-static3.nict.go.jp\/japanese\/wp-json\/wp\/v2\/event\/6789\/revisions\/6850"}],"wp:attachment":[{"href":"http:\/\/cinetjp-static3.nict.go.jp\/japanese\/wp-json\/wp\/v2\/media?parent=6789"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}