search

hisfog / SfMNeXt-Impl

[AAAI 2024] Official implementation of "SQLdepth: Generalizable Self-Supervised Fine-Structured Monocular Depth Estimation", and more.
MIT License
83 stars 12 forks source link

Confusion about some hyper-parameters #18

Open LiuJF1226 opened 10 months ago

LiuJF1226 commented 10 months ago

A nice work! but I am confused about some hyper-parameters.

LiuJF1226 commented 10 months ago

First is model_dim, it is set to 64 in resnet_320x1024.txt and resnet_192x640.txt, but it's 32 in args_res50_kitti_192x640_train.txt, 56 in args_cityscapes_train.txt. Then is min_depth and max_depth. In most config files, they are set to 0.001 and 80. But max_depth is set to 10 in args_cityscapes_train.txt and max_depth is set to 0.01 in resnet_320x1024.txt. And also, query_nums, dim_out and so on have different settings I wonder how these hyper-parameters can influence the final performance. Now I want to train on Cityscapes from scratch, should I follow the same setting in args_cityscapes_train.txt (max_depth=10 is somewhat strange to me, why not 80?)

hisfog commented 10 months ago

max_depth does not influence the scores (Because self-supervised training doesn't require absolute depth scale, and in evaluation, median-scaling is used for align depth_gts and depth predictions), but min_depth does. You can change model_dim, query_nums, dim_out if you like, but these hyper-parameters can influence the scores, model parameters, model speed, etc.

LiuJF1226 commented 10 months ago

Thank you for your reply. Actually I think the influence of min_depth is also slight if it is small enough, i.e., setting it to 0.1, 0.01 or 0.001 may have a slight gap.