-
c3d 8 conv layers + 5 pooling layers + 2 fc layers + softmax homogeneous：3x3x3 s1 conv throughout pool1：1x2x2 kernel size & stride，rest 2x2x2 fc dims：4096 C3D video descriptor：fc6 activations + L2-norm dropout: 论文里面没提，但是下面那个github的实现里面加了
-
3d-resnet r18和r34，后续文章探讨了更深的resnet，有人说效果奇差：https://github.com/kenshohara/3D-ResNets local cfg = { [10] = {{1, 1, 1, 1}, 512, basicblock}, [18] = {{2, 2, 2, 2}, 512, basicblock}, [34] = {{3, 4, 6, 3}, 512, basicblock}, [50] = {{3, 4, 6, 3}, 2048, bottleneck}, [101] = {{3, 4, 23, 3}, 2048, bottleneck}, [152] = {{3, 8, 36, 3}, 2048, bottleneck}, } identity shortcuts：use zero-padding 7x7x7 stem + 3x3x3 conv blocks + GAP + 400d-fc + softmax
-
pseudo-resnet 伪3d：通过1x3x3和3x1x1的S和T来实现 bottleneck 循环ABC bottleneck blocks model size：r50: 92M，p3d-r50: 98M input: 16x160x160 with an extra dropout layer with 0.9 dropout rate