depth loss and metric

[liyangliu]

Hi, @ozansener, I find that there may be some mismatch among the depth loss, metric and reported results in the paper. In the depth loss, those pixels with <0 loss are excluded, but in the depth metric, pixels with <250 target depth (normalized by mean and std) are all included, So I think they include all the pixels, maybe this is a bug because 250 should be used for segmentation but not depth regression. Also, in Tab. 4 of your paper w.r.t disparity error, are they the direct output of the depth branch of the model in your code? Thanks.

PS: For evaluating the depth regression performance, why should the gt and pred be cast as int?https://github.com/intel-isl/MultiObjectiveOptimization/blob/5d8e8343c56cf3184081f71cc4a661af64e7cf7e/multi_task/metrics.py#L46-L54