why can't predict occluding contours?
deng29 opened this issue · comments
Hello,
I load the model pre-trained on PBRS, then finetune on NYU. But when I run demo.py, the model can't predict occluding contours, the predicted result is blank. However, the other two task can work normally. I'm confused......
Hi,
Thanks for your interest in SharpNet. Could you share your test image? Also, if possible, can you share the trained weights? When finetuning on NYU, because they are not supervised anymore, the quality of contour decreases a bit.
However I am surprised that you cannot detect any contour anymore.
Hello,
I just use the official test split of NYU to test, the model training is completely in accordance with your instructions on github.
Hi,
Regarding occluding contour predictions
It sounds like the finetuning was done for too many epochs. It may be a mistake on my part regarding the number of epochs used for finetuning (80 seems quite high). I'll try to investigate it some time later.
In the meantime if you want to check results I suggest you use the final weights I had obtained.
Finally, if your priority is good occlusion boundary predictions, I would suggest to use the PBRS weights. Indeed as I mentioned, since they are not supervised during fine-tuning, their quality will progressively decrease. They only serve as guidance for the early stages of finetuning depth branch to keep the sharpness obtained with synthetic data.
Regarding other methods results
For Eigen, check here
For Laina, check here
For Fu (DORN), check here
For Jiao, I've put their predictions here.
Hi, let's discuss this offline, there are now several papers (with code, I hope?) that consider multi task supervision using semantic labels, but let's discuss this topic by email. Closing the issue for now