@JuewenPeng, Hi, this is a great work. I have some questions about the training: 1) During training, is the bn set to True? 2) What about share_weight? Is this set to True for both the models? 3) When training IUNet (while ARNet is frozen), is the adaptive resizing unused? If you consider adaptive resizing before ARNet for training IUNet, how do you manage different input sizes to IUNet? 4) Is the "connect_mode" set to "distinct_source" for both models? 5) Do you start training from scratch or the models are initialized by weights driven for another task? 6) You have mentioned that "To improve the generalization, we additionally augment the input disparity map with random gaussian blur, dilation and erosion." Could you please elaborate more on this? What are the maximum values for blur, dilation, and erosion? I would really appreciate it if you could help me with these questions.
@JuewenPeng, Hi, this is a great work. I have some questions about the training: 1) During training, is the bn set to True? 2) What about share_weight? Is this set to True for both the models? 3) When training IUNet (while ARNet is frozen), is the adaptive resizing unused? If you consider adaptive resizing before ARNet for training IUNet, how do you manage different input sizes to IUNet? 4) Is the "connect_mode" set to "distinct_source" for both models? 5) Do you start training from scratch or the models are initialized by weights driven for another task? 6) You have mentioned that "To improve the generalization, we additionally augment the input disparity map with random gaussian blur, dilation and erosion." Could you please elaborate more on this? What are the maximum values for blur, dilation, and erosion? I would really appreciate it if you could help me with these questions.