Fig. 2: Dimension rating task and results.
From: Revealing Key Dimensions Underlying the Recognition of Dynamic Human Actions
a N = 21 participants were asked to position 20 different reference action videos (e.g., ‘bicycling’) on each dimension. The dimension scale was separated into seven levels, ranging from very typical (left side) to very atypical (right side), based on the corresponding model weights of each dimension. Each level was represented by up to three representative videos. The example dimension scale depicts the dimension ‘food’. b The 20 reference videos were randomly sampled from all available 768 action videos. Left panel: similarity matrix based on the human dimension ratings from N = 21 participants, right panel: similarity matrix based on SPoSE model weights corresponding to each of the stimuli. c Pearson correlation between the two similarity matrices shown in panel (b) (Pearson’s r = 0.78; P < 0.001; randomization test; 95% CI, 0.71–0.83). For this figure, the original MiT video frames were replaced by images under a free-to-use license (Pexels: Askar Abayev, Center for Ageing Better, cottonbro studio, KampusProduction, Miriam Alonso, Roman Odintsov, vitalina, Yan Krukau; Pixabay: brenkee, Michelle_Raponi, PublicDomainPictures; Unsplash: DDP, Dipin Bhattarai, EqualStock, Frames for your Heart, Jovan Vasiljević, Julia Tsukurova, Mary West, Michael Kahn, National Cancer Institute, Rithwick. Pr, Sincerely Media).
