Clarification of "1K Resolution" in Table 3 – HexPlane-all* Evaluation

[sandokim]

Hello, thank you for sharing your excellent work!

I have a question regarding the resolution used for evaluation in Table 3 of "Neural 3D Video Synthesis from Multi-view Video".

From my understanding:

• The original resolution of the Plenoptic Video Dataset is 2028×2704 (2.7K).
• In 4DGS paper Table 3, most models (except HexPlane-all) were evaluated at half resolution (1352×1024).
• In 4DGS paper Table 3 and Hexplane paper Table 1, HexPlane-all appears to have been evaluated using results trained at "1K resolution."
• The original paper of the Plenoptic Video dataset (Neural 3D Video Synthesis from Multi-view Video) states: "We evaluate all the models at 1K resolution, and report the average of the result from every evaluated frame."
• Based on this, my understanding is that the "1K setting" refers to 1024×768, as 1K is commonly interpreted in this way.

I would like to clarify:

Is the "1K resolution" used for HexPlane-all in Table 3 actually 1024×768?

Did the authors of Neural 3D Video Synthesis from Multi-view Video or HexPlane clarify that their "1K" setting corresponds to 1024×768?

Thank you in advance for your clarification!

[sandokim]

One more question I’d like to ask is about the reported quantitative results for Im4D on the Neu3D dataset.

In Table 3 of the 4DGS paper, the quantitative results for Im4D on the Neu3D dataset appear to be identical to the values reported for the cut_beef scene in Table 2 of the Im4D paper. I’d like to confirm whether this is indeed the case.

• Im4D Table 2

• 4DGS Table 3

According to the Im4D authors’ response in the following GitHub issue, it seems that the quantitative evaluation of Im4D on the Neu3D dataset was conducted only on the cut_beef scene:

zju3dv/im4d#2

However, the 4DGS paper does not clearly state whether the reported Im4D results on the Neu3D dataset in Table 3 are averaged over all scenes (coffee_martini, cook_spinach, cut_roasted_beef, flame_salmon, flame_steak, sear_steak), or whether they simply reported the results from the cut_beef scene in the Im4D paper. This creates ambiguity in interpreting the quantitative comparison.

It would be very helpful if this point could be clarified.

[guanjunwu]

Dear sandokim,

Thanks for your valuable comments!

In the writing of this paper, we didn't check it and only borrow the data from the original paper of HexPlane and Im4D. we will fix it in the next arxiv version.

Best,
Guanjun

[sandokim]

I checked the official HexPlane code and found that it downsamples Neu3D's dataset to a resolution of 1024 x 768 (1K). :)

https://github.com/Caoang327/HexPlane/blob/main/hexplane/dataloader/neural_3D_dataset_NDC.py#L208-L211

Thanks for the help — this solves my issue, so I’m closing it now.

Best,
Hyeseong Kim