Hi,
I tried to reproduce the results on the Cityscapes test set. I trained the model with 16 GPUs, but use 1024x1856 crop size because of the limited memory.
With the split cv0, I got 86.69 on the val set and 84.86 on the test set. However, there is still a 0.5 gap with your submitted results. So I trained another model with split cv3, but got worse results, only 86.63 on the test set.
Compared the results of two evaluations, I found that the model trained with train+val can get improvements on almost all classes, but have a significant drop on some classes, e.g. -1.0 on "terrain", -3.1 on "truck", -3.6 on "bus", -4.3 on "train".
Could you help me to solve these problems?
Thanks!
Hi, I tried to reproduce the results on the Cityscapes test set. I trained the model with 16 GPUs, but use 1024x1856 crop size because of the limited memory. With the split cv0, I got 86.69 on the val set and 84.86 on the test set. However, there is still a 0.5 gap with your submitted results. So I trained another model with split cv3, but got worse results, only 86.63 on the test set. Compared the results of two evaluations, I found that the model trained with train+val can get improvements on almost all classes, but have a significant drop on some classes, e.g. -1.0 on "terrain", -3.1 on "truck", -3.6 on "bus", -4.3 on "train".
Could you help me to solve these problems? Thanks!