Local Ensemble is Wrong!

[AaronCIH]

Dear Authors,

I tried to use local ensemble strategy with implicit function.
However, I found there are wrong grid_sample results in the IFA.ifa_feat function.
Thus, the query_feats for local ensemble are also incorrect.

In addition, the original local ensemble strategy should rely on the mlp function for continuous transformation, but bilinear interpolation will be executed in your implicit function firstly. The difference formulates are shown as follow:
LIIF[1]: out = sum(si / s * mlp(qi, dqi))
Yours: out = mlp(sum(si / s * qi, dqi))

In this case, your local ensemble equals to the simple bilinear up sampling.

[1] Learning Continuous Image Representation with Local Implicit Image Function, Yinbo Chen, Sifei Liu, Xiaolong Wang, CVPR21 oral (https://arxiv.org/abs/2012.09161), code: https://github.com/yinboc/liif

[chichengfengxue]

Dear Authors,

I tried to use local ensemble strategy with implicit function. However, I found there are wrong grid_sample results in the IFA.ifa_feat function. Thus, the query_feats for local ensemble are also incorrect.

In addition, the original local ensemble strategy should rely on the mlp function for continuous transformation, but bilinear interpolation will be executed in your implicit function firstly. The difference formulates are shown as follow: LIIF[1]: out = sum(si / s * mlp(qi, dqi)) Yours: out = mlp(sum(si / s * qi, dqi))

In this case, your local ensemble equals to the simple bilinear up sampling.

[1] Learning Continuous Image Representation with Local Implicit Image Function, Yinbo Chen, Sifei Liu, Xiaolong Wang, CVPR21 oral (https://arxiv.org/abs/2012.09161), code: https://github.com/yinboc/liif

Hello, I have read the code. The part of decoder does not use bilinear interpolation, but nearest neighbor sampling. The code simply concat the features of each scale to the MLP.

[chichengfengxue]

Dear Authors,

I tried to use local ensemble strategy with implicit function. However, I found there are wrong grid_sample results in the IFA.ifa_feat function. Thus, the query_feats for local ensemble are also incorrect.

In addition, the original local ensemble strategy should rely on the mlp function for continuous transformation, but bilinear interpolation will be executed in your implicit function firstly. The difference formulates are shown as follow: LIIF[1]: out = sum(si / s * mlp(qi, dqi)) Yours: out = mlp(sum(si / s * qi, dqi))

In this case, your local ensemble equals to the simple bilinear up sampling.

[1] Learning Continuous Image Representation with Local Implicit Image Function, Yinbo Chen, Sifei Liu, Xiaolong Wang, CVPR21 oral (https://arxiv.org/abs/2012.09161), code: https://github.com/yinboc/liif

The author's code is wrong and x is not being added.