search

eragonruan / refinenet-image-segmentation

a tensorflow implement of refinenet ,RefineNet: Multi-Path Refinement Networks for High-Resolution Semantic Segmentation
130 stars 35 forks source link

advise for the code #11

Open sanggehouzhihoujue opened 6 years ago

sanggehouzhihoujue commented 6 years ago

The original code is written in matlab, and the network architectures are appended in the code package. It's my honor to read your code in tensorflow, but there are some problems in your code different from the orignal network architectures. With respect, I list these codes in below pictures. qq 20180212104032 qq 20180212104128 qq 20180212104215 I will be waiting for your reply.

sanggehouzhihoujue commented 6 years ago

@eragonruan

eragonruan commented 6 years ago

@sanggehouzhihoujue thanks for the sharing. For the first part about resnet101, I got this model from tensorflow/models. And I will re-check the second part soon.

shipengai commented 6 years ago

Hi,@sanggehouzhihoujue Did you test the model on PASCAL VOC validation set? and train it on your own dataset? Thanks. I'm looking forward your reply

sanggehouzhihoujue commented 6 years ago

@shipeng-uestc I trained it on the VOC-person part dataset, but the result is not so good. There should be some differences from the original code in matlab.

shipengai commented 6 years ago

Hi,@sanggehouzhihoujue . Thank you very much for your reply. Did you run the author's(@eragonruan ) demo.py? When I use the trained model from his link, some error about BN layer occurs. '''NotFoundError (see above for traceback): Key feature_fusion/Conv_37/BatchNorm/beta/ExponentialMovingAverage not found in checkpoint'''

sanggehouzhihoujue commented 6 years ago

@shipeng-uestc I ran the demo.py successfully. Do you create the dir named checkpoint?

shipengai commented 6 years ago

@sanggehouzhihoujue I think my ckpt has some error. Can you share me an old code or ckpt? Thanks.

sanggehouzhihoujue commented 6 years ago

@shipeng-uestc Didn't you train it on the VOC ? The ckpt is conducted at the training stage. The ckpt trained by myself can't result in good performance. What's more, I only trained on the part dataset.