Dear author, first of all, I would like to express my sincere gratitude for sharing your ideas on this work!
I have been trying to reproduce your work DiffVS based on the methods described in the paper recently. I have simply set up the overall architecture for now, but I am still unsure about some aspects of the training details and dataset processing. Therefore, I would like to ask for your advice.
There are mainly a few issues I'd like to confirm:
What specific rules were followed in the selection of the HEMIT test dataset mentioned in the paper? For instance, how were blank areas, low-signal areas, or overlapping areas removed? If it's convenient, could you provide the specific test set directory or sample list?
The original image size of HEMIT is 1024×1024. When training, is it directly resized to 512/256, or are small blocks cropped from the original image for training? If it is cropping, what are the specific cropping rules? For example, random cropping, center cropping, or sliding window cropping? What is the approximate stride?
Are the input dimensions in the verification and testing phases consistent with those in the training phase? Is the testing conducted on the cropped small images or on the original 1024×1024 images for inference?
These details will have some impact on the overall indicators of the reproduced results, so I sincerely hope to receive your reply. Thank you again for your work and sharing! I wish you all the best and everything goes smoothly.
Dear author, first of all, I would like to express my sincere gratitude for sharing your ideas on this work!
I have been trying to reproduce your work DiffVS based on the methods described in the paper recently. I have simply set up the overall architecture for now, but I am still unsure about some aspects of the training details and dataset processing. Therefore, I would like to ask for your advice.
There are mainly a few issues I'd like to confirm:
What specific rules were followed in the selection of the HEMIT test dataset mentioned in the paper? For instance, how were blank areas, low-signal areas, or overlapping areas removed? If it's convenient, could you provide the specific test set directory or sample list?
The original image size of HEMIT is 1024×1024. When training, is it directly resized to 512/256, or are small blocks cropped from the original image for training? If it is cropping, what are the specific cropping rules? For example, random cropping, center cropping, or sliding window cropping? What is the approximate stride?
Are the input dimensions in the verification and testing phases consistent with those in the training phase? Is the testing conducted on the cropped small images or on the original 1024×1024 images for inference?
These details will have some impact on the overall indicators of the reproduced results, so I sincerely hope to receive your reply. Thank you again for your work and sharing! I wish you all the best and everything goes smoothly.