some problem about trainning dataset

[futureflsl]

Afer I use labelme to get dataset.I can get img.png and label.png.But img.png bit depth is 24.I trained them and get one model,The test result is a whole black image.so I convert img.png to 8 bit depth.But I get low accuracy about 0.26 when I train.finally,The test result is a whole blue image.so I want to konw what bit depth did you use when you trained your data

[haixiansheng]

训练图像是灰度相机拍的8bit图像，label图像是24Bit的，要不然也不能多分类。