benjiebob / wldo Goto Github PK

Hey, I want to know how I can change your code and let them merely obtain 3D coordinates of keyjoints from dogs.

Good afternoon,

May I ask how do you install the neural renderer dependency? If I run regular pip install, it just fails for me. I have tried to fix some issues there and install it manually by copying the source code, and install it from the setup.py. The WLDO code then seems to run, but results are pretty bad. If I compare with the image from the README, I can see that I have worse results (with the pre-trained network) then it probably should be.

I am not 100% sure, but I guess that it might be the problem with the neural renderer dependency. Which version did you use and how did you install it?

Thank you.

I noticed that you have customized a dataset format and converted Animal Pose into this format. I'm curious about how you did it。Did you consider to provide the file like "changejsons.py"?

Hi Benjamin, your WLDO is excellent, I applied it on other animals and it can estimate their pose very good. However, I have a question about the keypoint you choose. For the dogs, I only see one keypoint on the whole torso, the tail-start. If there is another point around the neck, then it seems reasonable. However, there are only few points on the head. How do this one point define the direction of the torso? Or the tail-start point and the four points (point 4,2,5,11) can define the whole body. Thank you.

Hi, thanks for your great work. Are you going to release the training code?

Hi, it's me again, I got a question. This is source img:

and the result:

The shape looks very strange. I tried to modify "IMG_RES" from 224 to 480 and failed. How to get the same result in your paper? Thanks.

Hey, I ran "python demo.py" and got this error:

could you help me, thanks!

self.model_renderer = NeuralRenderer(
config.IMG_RES, proj_type=config.PROJECTION,
norm_f0=config.NORM_F0,
norm_f=config.NORM_F,
norm_z=config.NORM_Z)
synth_rgb, synth_silhouettes = self.model_renderer(
verts, faces, pred_camera)
synth_rgb = torch.clamp(synth_rgb, 0.0, 1.0)
TypeError: clamp(): argument 'input' (position 1) must be Tensor, not tuple

Is there a problem with the torch version? Or is the library now incompatible?

Issue reported by Le Jiang:

There are two functions "load_model_from_disk()" and "Model()"which are used in both demo and eval. One has "shape_family_id" and "load_from_disk" variables and one doesn't. The load_model_from_disk is defined by yourself so it can be different, but the Model() is the function from neural_renderer I think and it needs shape_family_id and load_from_disk. And I cannot run it now. Can you use the demo or do you have any update on this? Thank you so much!