Could we ignore global translation and only take joint rotation as input to MotionCLIP?

Hi. I just wonder if we can ignore global translation and only take joint rotation as input to your pretrained MotionCLIP? Does this make sense? Thank you very much!

Hi @VigorZero ! You can try switching this code:

MotionCLIP/src/models/modeltype/motionclip.py

Lines 204 to 206 in 4cc3e6d

    
           if is_amass and not self.align_pose_frontview:  # lose global orientation for amass dataset 
        
               print('NOTE: eliminating global orientation') 
        
               batch['output'][:, 0] = torch.tensor([1, 0, 0, 0, 1, 0]).unsqueeze(0).unsqueeze(2)

with:

        # FIXME - TEMPORARILY SWAP AXIS FOR VISUALIZATION
        batch['output_xyz'] = torch.cat([batch['output_xyz'][:, :, 1:, :], batch['output_xyz'][:, :, :1, :]], dim=-2)

This will look bad for text2motion but should work well for editing applications.

	if is_amass and not self.align_pose_frontview: # lose global orientation for amass dataset
	print('NOTE: eliminating global orientation')
	batch['output'][:, 0] = torch.tensor([1, 0, 0, 0, 1, 0]).unsqueeze(0).unsqueeze(2)