search

divamgupta / image-segmentation-keras

Implementation of Segnet, FCN, UNet , PSPNet and other models in Keras.
https://divamgupta.com/image-segmentation/2019/06/06/deep-learning-semantic-segmentation-keras.html
MIT License
2.92k stars 1.16k forks source link

I have high acc rate but the meanIU is too low,and the prediction result is bad. #169

Open buaacarzp opened 4 years ago

buaacarzp commented 4 years ago

I use the minidataset in image-segmentation-keras/test/example_dataset , my acc is:

2020-03-03 18:11:22.262469: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcudnn.so.7
512/512 [==============================] - 109s 212ms/step - loss: 0.3495 - acc: 0.8840
saved  ./checkpoints/fcn8s.model.0
Finished Epoch 0
Starting Epoch  1
Epoch 1/1
512/512 [==============================] - 101s 198ms/step - loss: 0.0812 - acc: 0.9695

and I use the picture in the train floder,but the result is too bad : image

I run the evaluate, I find the result is : 

{'frequency_weighted_IU': 0.05023619898653146, 'mean_IU': 0.02673938685619033, 'class_wise_IU': array([0.03116056, 0.06048602, 0.00145626, 0.05642503, 0.04454589,
       0.05107219, 0.00329684, 0.02061532, 0.02964784, 0.00604314,
       0.01148761, 0.00463594])}

what's wrong, please help me .

divamgupta commented 4 years ago

Hi,

Are you facing this issue with just fcn or all the models? Maybe try vgg-unet

buaacarzp commented 4 years ago

Hi,

Are you facing this issue with just fcn or all the models? Maybe try vgg-unet

The network I used is fcn8 inimage-segmentation-keras/keras_segmentation/models/fcn.py
. Today,I will read unet and pspnet, I hope there will be nothing wrong ....

buaacarzp commented 4 years ago

Hi,

Are you facing this issue with just fcn or all the models? Maybe try vgg-unet

Hi,bro, why your segnet use the Upsamping to decode? Tensorflow records the location of maximum pooling and use uppoolingto decode.

RookieXwc commented 3 years ago

I use the minidataset in image-segmentation-keras/test/example_dataset , my acc is:

2020-03-03 18:11:22.262469: I tensorflow/stream_executor/platform/default/dso_loader.cc:42] Successfully opened dynamic library libcudnn.so.7
512/512 [==============================] - 109s 212ms/step - loss: 0.3495 - acc: 0.8840
saved  ./checkpoints/fcn8s.model.0
Finished Epoch 0
Starting Epoch  1
Epoch 1/1
512/512 [==============================] - 101s 198ms/step - loss: 0.0812 - acc: 0.9695

and I use the picture in the train floder,but the result is too bad : image

I run the evaluate, I find the result is :

{'frequency_weighted_IU': 0.05023619898653146, 'mean_IU': 0.02673938685619033, 'class_wise_IU': array([0.03116056, 0.06048602, 0.00145626, 0.05642503, 0.04454589,
       0.05107219, 0.00329684, 0.02061532, 0.02964784, 0.00604314,
       0.01148761, 0.00463594])}

what's wrong, please help me .

Hello! I have the same problem, on my own data set, I have high acc rate but the meanIU is too low,and the prediction result is bad. How do you solve this problem?