the mIoU in Synthia dataset is only 28% about lsd-seg HOT 4 CLOSED

Let me get this right. For the complete synthia training, there are 19 classes. But following previous works and for fair comparison, we report results only on 16 classes. 3 classes (truck, train and terrain) are not reported. Please refer to our paper. Hence, your 28% actually amounts to ~33.5%. In our paper, we have reported the performance for 1024x512 sized images. For the size used commonly i.e. 640x320, our runs give 28.5-29.5%, which is close to what you are getting.

Hope this helps. Let me know.

from lsd-seg.

yes, you are right, thx!

==> Loading FCN8s_LSD model file: logs/MODEL-LSD_CFG-Adam_LR_0.00001000/model_best.pth.tar
==> Evaluating with CityScapes validation
tensorpack mIoU: 0.286795885361
tensorpack mean_accuracy: 0.385299141096
tensorpack accuracy: 0.798419604166
('Num classes', 19)
===>road: 78.47
===>sidewalk: 29.99
===>building: 75.9
===>wall: 6.9
===>fence: 0.85
===>pole: 21.73
===>light: 10.29
===>sign: 14.43
===>vegetation: 76.42
===>terrain: 0.0
===>sky: 77.73
===>person: 42.49
===>rider: 14.84
===>car: 62.78
===>truck: 0.0
===>bus: 11.48
===>train: 0.0
===>motocycle: 7.05
===>bicycle: 13.58
===> mIoU: 28.68

class16-mIoU= 34.06

from lsd-seg.

@dongzhuoyao. Hi, how did you manage to train it without mapped labels of SYNTHIA? Authors still have not provided this script :(

from lsd-seg.

using default setting, the mIoU is 28%.

Does anybody else get better result?

I think the training of GAN is very tricky.

Hi, I also used the default setting, but the mIoU is very low.

===>road: 0.33
===>sidewalk: 0.0
===>building: 76.71
===>wall: 0.36
===>fence: 0.91
===>pole: 0.08
===>light: 0.3
===>sign: 1.24
===>vegetation: 0.23
===>terrain: 0.01
===>sky: 0.0
===>person: 0.36
===>rider: 0.51
===>car: 0.0
===>truck: 0.02
===>bus: 0.04
===>train: 0.0
===>motocycle: 3.06
===>bicycle: 0.0
===> mIoU: 4.43

I don't know what went wrong, are there any parameters that need to be adjusted?
Looking forward to your reply

from lsd-seg.