For this I trained deeplabv3+ on pascal voc dataset(it have 21 classes).
I used a combined loss of dice loss and categorical crossentropy.
I just train this just for 9 epochs due to lack of time(before using this dataset I was using other datasets so I didn't had much time to train on this) ,so perform can be increased by training for longer.
here are some results: 1:
.png?raw=True)
_detail.png?raw=True)
2:
.png?raw=True)
_detail.png?raw=True)
3:
.png?raw=True)
_detail.png?raw=True)
4:
.png?raw=True)
_detail.png?raw=True)