Language Label Learning for Visual Concepts Discovered from Video Sequences

Abstract

Computational models of grounded language learning have been based on the premise that words and concepts are learned simultaneously. Given the mounting cognitive evidence for concept formation in infants, we argue that the availability of pre-lexical concepts (learned from image sequences) leads to considerable computational efficiency in word acquisition. Key to the process is a model of bottom-up visual attention in dynamic scenes. Background learning and foreground segmentation is used to generate robust tracking and detect occlusion events. Trajectories are clustered to obtain motion event concepts. The object concepts (image schemas) are abstracted from the combined appearance and motion data. The set of acquired concepts under visual attentive focus are then correlated with contemporaneous commentary to learn the grounded semantics of words and multi-word phrasal concatenations from the narrative. We demonstrate that even based on a mere half hour of video (of a scene involving many objects and activities), a number of rudimentary concepts can be discovered. When these concepts are associated with unedited English commentary, we find that several words emerge - approximately half the identified concepts from the video are associated with the correct concepts. Thus, the computational model reflects the beginning of language comprehension, based on attentional parsing of the visual data. Finally, the emergence of multi-word phrasal concatenations, a precursor to syntax, is observed where they are more salient referents than single words.