A Question About Input Resolution and depth prediction on nuScenes-mini

[zzh-yun]

Dear Authors,
Thank you for providing this excellent project! I have tested the model on the nuScenes-mini dataset and obtained the following results:

However, the performance metrics (especially Abs Rel and delta1 < 1.25) do not appear to be optimal. The depth visualization results are shown below for your reference:

I would like to ask for your guidance: does the input image resolution have an impact on the model's performance? My input image resolution is 896 × 1600, and I wonder if this resolution might be the reason for the suboptimal results mentioned above.
Looking forward to your valuable reply and guidance.

Best regards,

zihao

[zzh-yun]

Sorry, I forgot to normalize the images. The correct results are shown below, but they still don’t seem to reach the reported score of 0.953 on nuScenes in the paper.