Low test accuracy with released cityscapes_weights_before_bn_merge.caffemodel

[xurannlpr]

1 road: 0.3766 2 sidewalk: 0.0000 3 building: 0.0000 4 wall: 0.0000 5 fence: 0.0000 6 pole: 0.0001 7 traffic light: 0.0000 8 traffic sign: 0.0000 9 vegetation: 0.0000 10 terrain: 0.0000 11 sky: 0.0000 12 person: 0.0000 13 rider: 0.0000 14 car: 0.0000 15 truck: 0.0000 16 bus: 0.0000 17 train: 0.0000 18 motorcycle: 0.0000 19 bicycle: 0.0000 Mean IoU over 19 classes: 0.0198 Pixel-wise Accuracy: 37.66%

this is the test accuracy of the released cityscapes_weights_before_bn_merge.caffemodel on the cityscapes validation dataset. Does anybody know what's the reason behind this.

[jianbozhanghit]

@xurannlpr Hi, I met the same problem, could you please tell me how to solve it?