ConvCRF about the training curve

Hi @MarvinTeichmann , thanks a lot for your sharing valuable code!

I have a question about the learning curve, from the Fig. 4b, after equipped with the CRF module (epoch 100), the Train mIoU has a large jump, but seems to stay constant after that. Does this mean the CRF module learns nothing? Or maybe the network quickly converged from epoch 100 to epoch 101? Maybe I should check how the parameters of the CRF module changed from epoch 100 to epoch 250.

May 15 '18 17:05 glhfgg1024

hi @glhfgg1024,

thanks for getting in touch with me. My interpretation is that the model adapts very quickly to the added CRF. The train mIoU is definitely not constant after epoch 101. It only improves slower afterwards, but it can clearly be seen that the model still learns useful features. The val mIoU does not have such a jump and it climbs slowly from around 69% at epoch 100 to 72% at epoch 250.

May 15 '18 18:05 MarvinTeichmann

Understood. Thanks a lot for your comment and interpretation. From the Figure 2 of your paper, it is clear that the parameters can be learned and help improve the performance (maybe owing to good initialization for those parameters in CRF).

May 15 '18 18:05 glhfgg1024