ICCV Daily 2019 - Friday

need to pass the latent variable to the next step and then do it again and again. After several steps, the past information may vanish. Our method is a feed-forward method so that we don’t have that problem. Also, we’re trying to capture the dependencies between different joints using a fully connected network which means every joint is correlated to all the other joints, and that’s different from previous works which use a fully connected layer or convolutional layer to capture the dependencies between different joints. If you use a convolutional layer, then the dependencies will depend on the kernel size. Also, it depends on how you arrange your data. Our method is more flexible because it captures the dependencies between all the other joints.” Their work so far has been looking at predictions based on the human skeleton and stick-based human representation , but thinking about next steps, Wei points out that humans are 3D. We are tall, fat and skinny, and we interact with the world with our 11 Wei Mao DA I L Y skin, not our bones. That’s why in the future they will be trying to estimate human shape as well. In conclusion, Miaomiao tells us why she thinks this work is crucial for self-driving cars : “If we want cars to drive autonomously on the road , they need to understand not only the behaviour of other cars, but also the behaviour of humans. We need to predict what their future actions might be. For example, whether the human will cross the road or not. Colleaguesworking in the autonomous driving industry have confirmed that this is the next step that they will be working on , so it’s an important and interesting project for students.” To find out more about this work, come along to Wei’s oral and poster today at 14:52 in Hall D2. “If we want cars to drive autonomously on the road, they need to understand not only the behaviour of other cars, but also the behaviour of humans."