search

zhengyang-wang / Deeplab-v2--ResNet-101--Tensorflow

An (re-)implementation of DeepLab v2 (ResNet-101) in TensorFlow for semantic image segmentation on the PASCAL VOC 2012 dataset.
GNU General Public License v3.0
174 stars 90 forks source link

Results are similar to the original deeplab v2 paper? #5

Closed dyz-zju closed 6 years ago

dyz-zju commented 6 years ago

Hello, your implementation has achive the results compared to the original paper? I want to known,thanks!

zhengyang-wang commented 6 years ago

Compared to the original deeplabv2 paper, two things were not implemented:

  1. MSC: Employing mutli-scale inputs with max fusion.
  2. CRF

My mIoU on val is 75.1, which is reasonable.

dyz-zju commented 6 years ago

Hello, I just run your code on augmented pascalvoc datasets, and not change anything. After iter 20000, I test just get mIoU 68.8%, do you know why? Thanks!

dyz-zju commented 6 years ago

And more, the loss just shocks, after iter 20000, the final loss is aroud 1.3.

zhengyang-wang commented 6 years ago

Which pre-trained model did you use?

zhengyang-wang commented 6 years ago

And you may notice that you need to change 'valid_step' in main.py for testing/validation.

dyz-zju commented 6 years ago

I use deeplab_resnet_init.ckpt as my pertained model，and valid step is changed to 20000.

zhengyang-wang commented 6 years ago

Your final loss is correct. I cannot figure out what happened in your case without more information. I tested the code for several runs, the mIoU results for val are consistently around 75.

licaizi commented 6 years ago

@dyz-zju 您好，我也只跑出了0.68，请问后来您的结果怎么样了

zhengyang-wang commented 6 years ago

Please check https://github.com/zhengyang-wang/Deeplab-v2--ResNet-101--Tensorflow/issues/7

dyz-zju commented 6 years ago

@CaiziLee you need to use another dataset processed by author, the url can be found in somewhere.