WeiDong Chen
WeiDong Chen
看你错误的位置是你的tfrecord解析脚本中feature的定义有问题；请查看示例代码中tfrecord的生成脚本。label是一个one_hot类型的vector，被编码成了一个string，要使用decode_raw解码出来
Do you load pretrained regressor
you can use sigmoid instead linear, but as sigmoid suffers saturation problem, it may not be btter.
As the 2D keypoint loss calc. with camera scale, this may be does not matter.
When use offscreen render mode in Pyrender with platform='osmesa', it will produce 3 channel image, which means the alpha channel is lossed. I don't konw why, but it is. Instead...