Finetuning Resnet-50 didn't get sharp result about fcrn-depthprediction HOT 4 CLOSED

I cannot really answer with certainty regarding what goes wrong in your experiments, sorry. The whole setup is very different, including a re-implementation in caffe, fewer samples and the depth ground truth in log scale. All these could be possible issues.
Also the batch size sounds rather strange as this model shouldn't need that much space. I could easily fit a batch of 8 images in a 4GB GPU during training, for example.

from fcrn-depthprediction.

@iro-cp
Thanks for your reply. In the last two day, I did more experiments and got better result. As you said, using more data, or a more reliable loss function may achieve even better result.
About the memory, it's my mistake. It does not cost that much.
Thanks again.

from fcrn-depthprediction.

@xhwang @iro-cp Hi, facing the similar issue of blurred predictions while training the model(even train images output is blurred ... )

decoder- random initialisation
encoder - fcrn pre-trained weights
batchsize - 10
learning_rate - 0.005
momentum - 0.9
augmentation - same as fcrn

what did you change to get the better results.

from fcrn-depthprediction.

@Flying-Sailor Hi I am facing the same problem as you said. Have solved it?

from fcrn-depthprediction.