"warm-up" the computation in lower resolution.

[ywyue]

Hi authors, it was mentioned in the paper 7.1 Implementation - Optimization Details: "For stability, we "warm-up" the computation in lower resolution. Specifically, we start the optimization using 4 times smaller image resolution and we upsample twice after 250 and 500 iterations."

However, I didn't find the corresponding code - Could you please point it out to me?

Thanks in advance!

[grgkopanas]

The warm phase in low resolution seemed to helped in some experiment we originally did. Later on after fixing a few initialization issues here and there we tried to ablate it again and showed no impact so we removed it for the shake of simplicity.

…

On Mon, Oct 30, 2023, 16:49 Yuanwen Yue ***@***.***> wrote: Hi authors, it was mentioned in the paper 7.1 Implementation - Optimization Details: "For stability, we "warm-up" the computation in lower resolution. Specifically, we start the optimization using 4 times smaller image resolution and we upsample twice after 250 and 500 iterations." However, I didn't find the corresponding code - Could you please point it out to me? Thanks in advance! — Reply to this email directly, view it on GitHub <#412>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACGXXYMDAXS6SRNECVNW7J3YB7D6ZAVCNFSM6AAAAAA6WH4QE6VHI2DSMVQWIX3LMV43ASLTON2WKOZRHE3DQNRZGA3TEOI> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[ywyue]

Thank you for the reply! A quick follow-up question: I noticed that if the input image is larger than 1600px, it will be automatically resized to 1600px. However, the FovX and FovY still come from the original camera info (here) calculated from the original image width/height. I wonder is there any motivation behind this?

[jb-ye]

@grgkopanas What if I train the image on 4k+ images without downscaling to 1600, can coarser-to-fine warm-up help for high resolution input images?

[grgkopanas]

Potentially, many of these tricks are hand wavy and experimentally based so you can only be sure if you try

…

On Wed, Nov 1, 2023 at 8:28 PM J.Y. ***@***.***> wrote: @grgkopanas <https://github.com/grgkopanas> What if I train the image on 4k+ images without downscaling to 1600, can coarser-to-fine warm-up help for high resolution input images? — Reply to this email directly, view it on GitHub <#412 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACGXXYPXTIYLJ2U5EERHFRTYCKPFRAVCNFSM6AAAAAA6WH4QE6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTOOBZGUZTCMBUGU> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[kwea123]

@ywyue fov doesn't change if you maintain the image aspect ratio while resizing

[chenqi13814529300]

@grgkopanas What if I train the image on 4k+ images without downscaling to 1600, can coarser-to-fine warm-up help for high resolution input images?

Have you tried this please, I suddenly see this problem?