Poor performance on landmark detection demo case

[tearscoco]

I've set up the env correctly according to the doc. I did get an [The model and loaded state dict do not match exactly] error during demo test but I think it's ok as I loaded the model successfully and got the result (using the resnet ckpt from model zoo).

However, the detection performance on the demo image seems poor as shown below. The same poor performance happened to my own data as well.

Could you please help me have a look at it? Thanks a lot!