What would be the optimal image size for testing the LoFTR model for pose estimation?

[Coronal-Halo]

I see in the paper the authors say they trained LoFTR on Scannet with images resized to 480 x 640 pixels, and on Megadepth with images resized to have the longer dimension equal to 840 (and 1200 for validation). However, I wonder how the image sizes will affect the training/testing results? If the original image is bigger, will resizing it to a smaller dimension increase/decrease the accuracy? I personally find in my case, for pose estimation, resizing the image to have the longer side = 840 pixels decreases the testing result compared to resizing to 1200 since many of my testing images have dimensions bigger than 1000 x 1000, I don't know if this problem is investigated.

[angshine]

Generally, larger image resolution leads to better accuracy. Therefore, I recommend keeping your original resolution for matching regardless of the training resolutions. One exception is when the images to be matched differ a lot in scale. I would recommend aligning their scales through image resizing before matching.

[Coronal-Halo]

I see, thank you for the prompt reply. I have another question, why did you resize the images to 840 pixels for training but 1200 for validation on Megadepth?

[shy922]

How does this data set download too big for me？