Cross-modal Representation Learning for Understanding Manufacturing Procedure

Assembling, biochemical experiments, and cooking are representatives that create a new value from multiple materials through multiple processes. If a machine can computationally understand such manufacturing tasks, we will have various options of human-machine collaboration on those tasks, from video scene retrieval to robots that act for on behalf of humans. As one form of such understanding, this paper introduces a series of our studies that aim to associate visual observation of the processes and the procedural texts that instruct such processes. In those studies, captioning is the key task, where input is image sequence or video clips and our methods are still state-of-the-arts. Through the explanation of such techniques, we overview machine learning technologies that deal with the contextual information of manufacturing tasks.