scripts/README.md

Important hyperparameter explanations

MVDataset

n_all: No. of data samples uniformly chosen from the whole h5 list. num_views: No. of input views + novel views num_render_views: No. of novel views (then we will have num_input_views = num_views - num_render_views) random_order: whether data augmentation of random shuffling all images happens. random_nv_nr: a list of pairs (num_views, num_render_views). this will lead to mixed length training. (for MVDp+'s stage 2) n_ref: number of reference views (and MV-DUSt3R+ will take them for multiple paths)

Model

pts_head_config: skip: whether there is a skip layer in the end of linear layer for higher frequency short path. m_ref_flag: whether it's MV-DUSt3R or MV-DUSt3R+