Bad prediction for multiple classes

[baniani]

Hi WyattAutomation@, thank you for putting together a really useful tutorial.

I've been trying to train a model with three classes: Human body, head, and hand and here is the structure and settings:

obj.names: Human_body Human_head Human_hand

obj.data: classes= 3 train = /floyd/home/dataset/p1_dataset/train.txt valid = /floyd/home/dataset/p1_dataset/test.txt names = /floyd/home/darknet/obj.names backup = /floyd/home/darknet/backup

cfg file: [net] batch=32 subdivisions=8 width=416 height=416 channels=3 momentum=0.9 decay=0.0005 angle=0 saturation = 1.5 exposure = 1.5 hue=.1 ... filters=24 ...

train.txt (15000 images): /floyd/home/dataset/p1_dataset/Human_body/f62eea2d18ea0ca2.jpg /floyd/home/dataset/p1_dataset/Human_body/02eb772721398fc5.jpg /floyd/home/dataset/p1_dataset/Human_hand/b4711678a4f11dec.jpg /floyd/home/dataset/p1_dataset/Human_head/457522f52a4b3643.jpg

The images and their order is random.

Each label file has only one type of class in it, for example, 0000bee39176697a.txt:

0 0.36531250000000004 0.5911895 0.275625 0.814097 0 0.1825 0.6462555 0.1875 0.689867 0 0.721875 0.517621 0.5075 0.9629960000000001

Is this structure ok? Am I missing something? The problem is even after 5k, darknet can't detect anything.