if using ordinary Conv2D, or drawing 17 joints in one heatmap, is OK?

[liuqunzhong]

1、Do you test the performance if using ordinary Conv2D, not using Conv3D? 您好，请问，在一些NPU上不支持3D卷积，如果不使用3D卷积，而是使用普通的2D卷积，效果如何呢？

2、Do you test If drawing 17 joints or limbs of one frame in one heatmap?The input size is from K × T × H × W to T × H × W。 请问，如果将某帧的点和肢体都画在一幅图上，是否可行呢？像论文中的为了便于展示将点画在一张图中的Joint stream Input、Limb stream Input图片。

[kennymckormick]

1. PoseConv3D uses 3D conv for temporal modeling, if replacing it with normal 2D conv, the temporal modeling capability may significantly deteriorate and the recognition accuracy will drop.
2. It's feasible, but again, it may lead to some performance drop. A better alternative might be reshape the input from K x T x H x W as (K * T) x H x W. You will still get a 2D input but more information can be kept.