The objective of this project was to try to accurately anticipate the sequence of actions that will be carried out in the next five minutes to perform a certain task. The computer scientists considered that the kitchen tasks were perfect to make the tests, and prepared their program so that it could learn to anticipate the actions. In order to make accurate predictions, first, it is necessary to train the software. For this, the team used 40 videos of salads preparation. All of them lasted approximately 6 minutes and contained an average of 20 different actions. With this training, the algorithm learned what are the usual actions that are carried out to prepare a salad, as well as the duration of each of them. Later they continued adding videos so that the system learned to anticipate the actions. “We want to predict the time and duration of the activities minutes and even hours before they happen,” says Jurgen Gall, the director of the research. With the data obtained through this material, the software learned to identify the first actions of the video and then predict what the chef would do next. However, later, the engineers evaluated their ability to anticipate both the sequence and the duration of the tasks.
“The accuracy was over 40 percent for short forecasting periods, but then declined the more the algorithm needed to look into the future,” says Gall. “For actions that lasted more than 3 minutes, the program succeeded in 15% of cases.” This study is only a first step in the field of data prediction, so although the initial results have been promising, there is still a long way to go. So, what do you think about this? Simply share all your views and thoughts in the comment section below.
