In the case where images with a LABEL_RATIO>0 are sparse (something I've encountered a lot with power infrastructure data), the program spends a lot of time writing blank tiles that are later thrown out when making sample_filtered. Additionally, in this case, checking a label for equality against the blank image file is much faster than creating an entry in label-counts.txt.
In the case where images with a LABEL_RATIO>0 are sparse (something I've encountered a lot with power infrastructure data), the program spends a lot of time writing blank tiles that are later thrown out when making sample_filtered. Additionally, in this case, checking a label for equality against the blank image file is much faster than creating an entry in label-counts.txt.