I tried to train SDFA-Net by myself, but I find that the training process seems to be unstable. For example, I find that two consecutive epochs may show very different results (like abs_rel=0.100 and abs_rel=0.120). This phenomenon also brings troubles for me to reproduce your results reported in the paper.
I also tried to train SDFA-Net but disable the crop & resize data augmentation (the image size is set to 192 x 640 during both training and test.), and the training process seems to be more smooth, e.g., large performance change will not happen between two consecutive epochs. So I doubt that the stability of the training process may be related to this data augmentation.
But lower resolution seems to harm the final performance, so I still failed to reproduce your results.
Do you have any idea or have tried something about this?
Hi, thank you for open-sourcing your work!
I tried to train SDFA-Net by myself, but I find that the training process seems to be unstable. For example, I find that two consecutive epochs may show very different results (like abs_rel=0.100 and abs_rel=0.120). This phenomenon also brings troubles for me to reproduce your results reported in the paper.
I also tried to train SDFA-Net but disable the crop & resize data augmentation (the image size is set to 192 x 640 during both training and test.), and the training process seems to be more smooth, e.g., large performance change will not happen between two consecutive epochs. So I doubt that the stability of the training process may be related to this data augmentation. But lower resolution seems to harm the final performance, so I still failed to reproduce your results.
Do you have any idea or have tried something about this?
Thanks!