LearningSpatiotemporalFeatureswith3DConvolutionalNetworks（2014）

文章引用格式：
Tran D, Bourdev L, Fergus R, et al. Learning spatiotemporal features with 3d convolutional networks[C]//Proceedings of the IEEE international conference on computer vision. 2015: 4489-4497.

大概介绍：

kernel size为3×3×3.

实际模型阶段：

Training is done on the Sports-1M train split. As Sports-1M has many long videos, we randomly extract five2-secondlongclipsfromeverytrainingvideo. Clipsare resized to have a frame size of 128×171. On training, we randomlycropinputclipsinto16×112×112cropsforspatial and temporal jittering. We also horizontally flip them with 50% probability. Training is done by SGD with minibatch size of 30 examples. Initial learning rate is 0.003, andisdividedby 2 every150Kiterations. Theoptimization is stopped at 1.9M iterations (about 13 epochs). Beside the C3Dnettrainedfromscratch,wealsoexperimentwithC3D net fine-tuned from the model pre-trained on I380K