Ziwei He¹, Yen-Ju Chen², Yung-Hao Yang¹, Shin'ya Nishida¹; ¹Graduate School of Informatics, Kyoto University, Kyoto, Japan, ²Department of Psychology, Soochow University, Taipei, Taiwan

Humans perceive motion in a highly flexible and hierarchical manner. For example, when observing a walking person, we simultaneously perceive the global translation of the body and the local oscillatory movements of the limbs. Classical work shows that the visual system can infer multiple latent motion components and their hierarchical organization. More recent work has shown that observers can dynamically assign motion to infer reference objects from dot motion. However, the perceptual limits of hierarchical motion decomposition remain unclear. Here, we investigated these limits using a vector matching paradigm. The stimulus consisted of four hierarchical levels: a static background, a large circle, a middle circle, and a small circle. Each layer was marked by a distinct color and composed of randomly distributed dots moving at a constant speed and direction. The motion of each deeper layer was defined as the vector sum of the motions of all preceding layers, thereby forming a nested structure. Before stimulus onset, participants were informed of the reference and target layers. The stimulus was presented for 1 s and repeated three times, after which participants reproduced the perceived relative motion of the target by adjusting the direction and speed of matching disks. Performance was evaluated by comparing the reproduced motion vectors with both the ground-truth absolute motion and the intended relative motion (target minus reference) across different hierarchical configurations. Accuracy was highest when the reference and target formed a direct hierarchical connection ({reference{target}}). Performance declined for both standard induced-motion configurations ({reference{irrelevant{target}}}) and double induced-motion configurations ({reference{irrelevant1{irrelevant2{target}}}}). In contrast, task-irrelevant child ({reference{target{irrelevant}}}) and grandchild ({reference{target{irrelevant1{irrelevant2}}}}) layers had minimal influence on performance. These findings demonstrate that, although humans can decompose hierarchical relative motion, their precision is systematically constrained by hierarchical distance, with dominant contributions from higher-level reference objects and little contamination from lower-level, task-irrelevant motion.

Acknowledgements: This study is supported by JST SPRING JPMJSP2110, Kakenhi JP24H00721