The first predict frame problem about auto_conditioned_rnn_motion HOT 6 OPEN

The first frame in the bvh file should always be a T pose. It seems that your rendering software didn't render all the joints. Have you try to open the .bvh file in MotionBuilder?

from auto_conditioned_rnn_motion.

Thank you for your answer~

Yes,you are right.The first frame in the bvh file is always a T pose.I thought it is not the first predict frame, sorry. But the second frame in the bvh file is always that pose as the picture shows, I don't know how to solve it,and I I also don't know what caused this problem.Please help me.

Thank you~~

from auto_conditioned_rnn_motion.

I can't see the problem with only your description. Can you send a download link of your bvh files?

from auto_conditioned_rnn_motion.

prediction_out_02.bvh.tar.gz

Here attached a sequence of motion prediction that lasts about 415 frames before freezing and the weights of the model were trained on indian dance. We opened the .bvh file using Blender 2.79b. The second frame of the generated animation figure appears weird pose same with picture above.

from auto_conditioned_rnn_motion.

I watched your motion sequence. It seems to be a normal result. The reason that the first generated frame (here is frame two) is not good is that the initial memory (c and h) is set to be zeros, which is not at the manifold of the memories that can generate good poses. And because the initial memory is always set to be zeros during training, the output of the first frame is likely to be similar.

from auto_conditioned_rnn_motion.

Thanks for your replying.
You are right, the initial state was set to 0 so that the model can't predict well. As the memory has been updated, everything goes well.
We are trying to feed as less as possible ground truth as inputs to our model before it starts to rely on its own previous prediction to generate a new one. So the first few frames are important.
Thanks anyway, we will look into more details.

from auto_conditioned_rnn_motion.