search

rongakowang / DenseMutualAttention

[WACV 2023] Interacting Hand-Object Pose Estimation via Dense Mutual Attention
MIT License
36 stars 3 forks source link

Pretrained checkpoints (HO-3D & DexYCB) do not work #5

Closed taeyunwoo closed 1 year ago

taeyunwoo commented 1 year ago

Hello, thank you for your works.

I tried to follow up your code with the uploaded pretrained checkpoint of models (HO-3D & DexYCB). However, I think the checkpoint does not work in my environment.

Actually, I commented out the neural_renderer part in here. (Also I erase which are related with it.) However, I think it does not affect the forward of the module, if not I apologize for my fault.

As I checked the checkpoints, the 'epoch' of the trained checkpoint is 12, while the paper said you trained 25 epochs. Also, when I visualize the network output, the hand and object are not reconstructed well. Significantly, the hand does not look the hand, and the object coordinates far from the hand.

Here is the visualizing result in DexYCB. image

Thank you for your replying.

rongakowang commented 1 year ago

Hi,

Can you please first confirm that you can reproduce the Table 2 results using the weight

The visualization looks to me more like that you are not connecting vertices by the correct faces

taeyunwoo commented 1 year ago

In my settings, MJE / MCE / PD = 8.955 / 364.84 / 0.0. Also, the mesh faces are not wrong, as in vertices visualization below.

image

I think, something's wrong in my setting. Again, I re-examine my setting thoroughly.