Can the Hi-SAM be trained to segment the watermark, such as some lines? 