About the Freihand demo about i2l-meshnet_release HOT 16 CLOSED

You used a body network.. please read the README carefully and use a hand network.

from i2l-meshnet_release.

seems like you did right!

from i2l-meshnet_release.

I used camera parameters from FreiHAND annotations, which is available in the evaluation setting.
The root depth is obtained from RootNet, my ICCV19 paper link. You can download the depth values from here.

from i2l-meshnet_release.

What if I want to run on other images not from FreiHand? And can you share the pretrained model of RootNet?

from i2l-meshnet_release.

Sorry I didn't save the pre-trained RootNet on FreiHAND.
If you do not know the intrinsics, you can just use normalized intrinsics (fx=1500, fy=1500, cx=img_width/2, cy = img_height/2).
Or, if you assume a weak-perspective, you don't need the focal lengths because they are eliminated during camera-backprojeciton.

from i2l-meshnet_release.

I got a bad result by using (fx=1500, fy=1500, cx=img_width/2, cy = img_height/2). (fx=500, fy=500, cx=img_width/2, cy = img_height/2) is much better.

from i2l-meshnet_release.

good. you can use a weak-perspective camera model to eliminate the focal lengths.

from i2l-meshnet_release.

@mks0601 I test more hand pictures, but they seem not good. Can you give some suggestions?

from i2l-meshnet_release.

You'd better visualize mesh_out_img at here. This 2.5D representation does not require intrinsics and root depth

from i2l-meshnet_release.

@mks0601 I used function "vis_mesh" to visualize the x,y. But It's still not good.

from i2l-meshnet_release.

I can't give you more detailed guidance because I don't know your testing images

from i2l-meshnet_release.

@mks0601
I test on FreiHand dataset with image like this

And got this result

How can I improve result on Hand dataset??

from i2l-meshnet_release.

@mks0601 Sorry but your mean is using checkpoint snapshot_24.pth.tar (of FreiHAND), right?
Because when I used it and ran demo with command <python demo.py --gpu 1 --stage param --test_epoch 24>, it got error size mismatch for module.human_model_layer.th_faces, module.human_model_layer.th_weights

from i2l-meshnet_release.

You'd better just run a full code, not a demo code if you want to run the hand model. There should be several modifications for the hand demo code. If you want to run the demo code, could you set

trainset_3d = ['FreiHAND'] # MuCo, Human36M, FreiHAND
trainset_2d = [] # MSCOCO

in the main/config.py?

from i2l-meshnet_release.

@mks0601 Surely, I have modified this config file. test_set I set for FreiHAND too :( . I am so confused about this problem

from i2l-meshnet_release.

@mks0601 . After investigating many times. I get this results. Did I do right??

from i2l-meshnet_release.