ICCV Daily 2019 - Tuesday

applicability of the transfer learning.” Okan says the biggest takeaway of this task is that people should select or make their architectural design according to the given task . Looking around the conference this year, he has seen that many works are using ResNet-50 or ResNet-101 and they do not check for any other architecture. He doesn’t think that is the right way. For example, in his paper you can see that the motion patterns are captured better with depthwise convolution. He tells us that MobileNetV2 has much more depthwise convolution compared to the other architecture and this is why it achieves the best performance on Jester dataset on MobileNetV2. He adds that it is better than ResNet-101 although it has 25 times less parameters and floating- point operations . The codes and pre-trained models used in this work are available for everyone to use. "Kinetics-600 can measure the capacity of the experimented networks. " 11 Okan Köpüklü DA I L Y M e e t your colleagues and find new friends at the reception tonight! Food and drinks will be prepared at the 3F lobbies of Auditorium, Hall D and Hall E, and the inside of Hall E. The reception will start at 18:00. Don’t be late!