[deeplab] Train DeepLabV3+ with NYU-Depth V2 dataset

Hi all,

I just tried to use NYU-Depth V2 dataset to train DeepLabV3+. But the loss value does not converge to a lower value. Here is the training settings:

• Input images manually shrunk to a resolution with 512x512
• Train images: 796, eval images: 655
• 14 classes (13 classes, id 1-13 + 1 background, id = 0)
• ignore_label = 0

1. Using ADE20K base model, final loss is around 0.4 - 0.5, overall mIoU: 0.6609779

   • train_crop_size 256x256, bs=8, last_layers_contain_logits_only = true, initialize_last_layer = false
   • eval_crop_size 513x513

2. Using ADE20K base model, final loss is around 0.6, overall mIoU: 0.620792508

   • train_crop_size 256x256, bs=8, last_layers_contain_logits_only = false, initialize_last_layer = false
   • eval_crop_size 513x513

3. Using PASCAL VOC base model, final loss is around 0.8, overall mIoU: 0.525445759

   • train_crop_size 256x256, bs=8, last_layers_contain_logits_only = true, initialize_last_layer = false
   • eval_crop_size 513x513

4. Using PASCAL VOC base model, final loss is around 0.8, overall mIoU: 0.551259339

   • train_crop_size 256x256, bs=8, last_layers_contain_logits_only = false, initialize_last_layer = false
   • eval_crop_size 513x513

Are these results acceptable ?

If I want to increase the mIoU value for the training with NYU-Depth V2 dataset, what parameters should I change and test ?

• Learning rate ?
• Momentum ?
• Weights ?

And as I am a newbie in deep learning field, I do not know the exact usage of above parameters and where I could change the "weight" of the model.

Please help.

Thanks.