Hyperparameters of several finetuning experiments
Opened this issue · 2 comments
Thank you for your great work!
Could you give the hyperparameters (like learning rate, batch size, crop size, and step number) that can reproduce the paper data in three finetuning experiments (KITTI, ETH3D and Middlebury), which are not explicitly described in the paper "RAFT-Stereo: Multilevel Recurrent Field Transforms for Stereo Matching".
Thank you very much.
These are the settings:
Middlebury: lr=2e-5, batch_size 2, crop_size=384x1000, steps=4k
ETH3D: lr=1e-5, batch_size 2, crop_size=384x1000, steps=4k saturation_range=[0 0] img_gamma=[0.5 2.0 0.8 1.2]
KITTI: lr=1e-5, batch_size 3, crop_size=320x1000, steps=5k
Thank you for your response above.
Can you also please specify the number of disparity field updates configured for fine-tuning and inference, for all three datasets? Thank you.