image_fitting

[Sugiiiii]

The current image_fitting does not seem to be the latest version. I encountered many variable mismatch issues while trying to run image_fitting.py. I would really appreciate it if you could upload the latest runnable version.

[Sugiiiii]

This is the result I got:

[yhc-1]

Same issue, could you please update your code?

[moonbow721]

It would be better if the codes for quantitative evaluation could be updated.

[moonbow721]

The implementation of image_fitting is quite different from the descriptions in the paper. For example, the bending
term for elbows and knees is not used here. Besides, why does the 'temp' term for motion sequences also appear here?

PoseNDF/experiments/image_fitting.py

Lines 179 to 213 in 2a9e4f1

	optimizer = torch.optim.Adam([smpl_init.body_pose], 0.02, betas=(0.9, 0.999))
	# Get loss_weights
	weight_dict = self.get_loss_weights()
	for it in range(iterations):
	loop = tqdm(range(steps_per_iter))
	loop.set_description('Optimizing SMPL poses')
	for i in loop:
	optimizer.zero_grad()
	loss_dict = dict()
	# convert pose to quaternion and predict distance
	pose_quat = axis_angle_to_quaternion(smpl_init.body_pose.view(-1, 23, 3)[:, :21])
	loss_dict['pose_pr']= torch.mean(self.pose_prior(pose_quat, train=False)['dist_pred'])
	# calculate temporal loss between mesh vertices
	smpl_init = self.body_model(betas=self.betas, pose_body=smpl_init.body_pose)
	# smpl_opt = self.smpl(betas=self.betas, pose_body=smpl_init.body_pose)
	temp_term = smpl_init.vertices[:-1] - smpl_init.vertices[1:]
	loss_dict['temp'] = torch.mean(torch.sqrt(torch.sum(temp_term*temp_term, dim=2)))
	# calculate data term from inital noisy pose
	if it > 0: #for nans
	data_term = smpl_init.Jtr - init_joints
	loss_dict['data'] = torch.mean(torch.sqrt(torch.sum(data_term*data_term, dim=2)))
	# Get total loss for backward pass
	tot_loss = self.backward_step(loss_dict, weight_dict, it)
	tot_loss.backward()
	optimizer.step()
	l_str = 'Iter: {}'.format(i)
	for k in loss_dict:
	l_str += ', {}: {:0.8f}'.format(k, weight_dict[k](loss_dict[k], it).mean().item())
	loop.set_description(l_str)

[HonglinChu]

where define self.loss_weight?

[HonglinChu]

I can not run image_fitting, it seems that the code have many errors

[HonglinChu]

where the self.loss_weight is defined?

[tacalvin]

Anyone manage to get this to work?