Which MVimgNet classes to intersect with Imagenet dataset classes?

[jesimonbarreto]

Hello,
Thank you for making the dataset available. I tried to find the same classes in both datasets using the list available but the number is lower than what was reported in the text.

"The objects in our MVImgNet are found or
used in human daily life (human-centric), where 65 classes
overlap with ImageNet."

Could you make it clear what these classes are in MVImgnet. Also, if possible, make them available in a folder separately for download, please?

Thank you.

[mutianxu]

Thanks for your question.

In fact, we provide "MVImgNet_by_categories" just for partial debugging/testing on specific categories, so it is incomplete. Please also note that splitting the classes is highly nontrivial, so currently, we can only provide limited categories.

To access all categories (65 classes, detailed in the supplementary material of the paper), you need to download all the mvi_xxx.zip files directly.

We have provided the Linux download script in README, so it will be very convenient now.

Hope this is helpful!

[jesimonbarreto]

Thanks for the answer. I looked at the supplementary material but couldn't find the 65 categories exactly. Could you tell me exactly where this detail is in the paper, please?

Thank you so much!

[mutianxu]

Sorry for the confusion!

So, we actually manually compare our categories with ImageNet categories.

Here is the category list of MVImgNet, you can compare it with ImageNet by yourself.

mvimgnet_category.txt

Thanks!

[jesimonbarreto]

I compared the class names, but the total is less than 65 classes.

Is there any detail beyond direct comparison of names?

Thank you for your attention and answers!

[matheusfreitasmv]

Hello,

I have the same question. I couldn't find the 65 classes that overlap with ImageNet.

Is it possible to provide these overlapped classes, please?

Thank you.

[mutianxu]

Hi,

Sorry, we actually manually compare each class (very burdensome and non-trivial), it does not require "literally same" classes, but just consider very similar classes that may denote the same object.

Therefore, a very simple method is to send the category text of both MVImgNet and ImageNet to CLIP, and calculate the similarities in the latent space.

These are all the suggestions we can provide at this time.

Hope this is helpful to you!