Error on depth estimation

[nicetryzz]

I have test the KITTI-MOT 0006 Reconstruction config. The depth estimation is similar with the Wandb log provided. The estimated depth is different from the mono depth.

[JiantengChen]

Hi! @nicetryzz

Thanks for your reply. Issue #92 might help. FYI.

[nicetryzz]

Thanks for your quick replay. I think utilize the sky weight on depth loss calculation is a good idea.
By the way, I have two more questions:

1. Does it reasonable to select rays from different images to calculate mono loss? Different image mono depth esitimate might have different scale and shift.
2. I notice that scale_factor is set to 0.1. Also while calculate monosdf_depth_loss, a scale and shift is applied on termination_depth. Does it necessary？

[wuzirui]

Hi @nicetryzz, for your questions:

1. it is true that "select rays from different images to compute scale and shift" is not making too much sense for the monodepth, but overall the scale and shift is only an approximation to match monodepth with LiDAR depths. For the practical sake, this is enough because monodepths are mostly consistent in a sequence.
2. I didn't get your questions sry. Could you elaborate more?