Different results when I train my own model & bugs

[ChlaegerIO]

Thank you for the nice work. The evaluation and demo with your weights and configuration works like in the paper. I have installed the library versions provided in the requirements.txt

My problem now is that I want to adapt, e.g. shrink, it to fit my purpose therefore I have to train my own models. So I tried to train the entire DELTAR model as it is provided by you. I use 25k nyu images. I have tweaked the code here and there to fit my data and I changed it with the two additional remarks from below. In this setting I trained it for 50 epochs, but I have also trained for 25 epochs with the settings provided in the configs for nyu. I then get the following results for validation:

and training loss:

images from the training set looks good though:

but it does not generalize that well:

My questions now are:

• How do you load/preprocess the nyu dataset?
  I used this python version of the toolbox: https://github.com/Lucass97/nyuv2_raw_dataset_extractor
• Do you have other suggestion what to watch out for?

Additional remarks:

• I found that the pointNet parameters where not trained in the network before as they are not added to the parameters for the optimizer in

    def get_10x_lr_params(self):  # lr learning rate
        modules = [self.decoder, self.depth_head, self.conv_out]

• In the nyu dataloader in "train" mode the image should be normalized t0 [0,1] (image = np.array(image, dtype=np.float32) / 255.0) before the random_crop(...) and train_preprocess(...) augmentation as it is clipped there to [0,1]