Hason Xia

@xindongzhang 你试过SBR训练regression网络吗？

是在计算loss的时候把这些点忽略掉吗

请问一下，为什么用您发布的hourglass网络在MPII上随机初始化训练时，loss很低

因为很低，验证集上loss没有变化趋势

> but if it's just one or two pixels off the border, we will still have a little bit gaussian remaining for that joint right? because the gaussian patch itself...