Question about the visualization of the result

[Ajeosi]

Thank you for your code for segmentation of CamVid data.
I ran the code without any change (10 epochs, Adam optimizer).

My question is,
How can I visualize the predicted image with 12 classes with 12 colors well?
All I get is the strange image with multiple horizontal lines.

When the imgs_mask_test.npy is made after running the code, it is (233,172800, 12) array.
To visualize the result of 1st sample of test set, , I reshaped this to (360,480,12) and,
np.argmax(reshaped image, axis=-1) was done to make each pixcel of (360,480) to represent the each class whose probability was highest.
When I plot the image np.argmax(~~~) by pyplot.imshow, I get the strange image with multiple horizontal lines.

Is there any good way to make a picture made of 12 colors representing each object, without horizontal lines?

(original image of test set)

(annotated image of test set)

(result with horizontal lines by prediction)

Thank you.

[SteveIb]

Hi @Ajeosi , did you find a solution for that? I'm facing the same problem with 3 classes.

[YJonmo]

Any solution?

[shabtayor]

I was able to solve this issue by changing the Reshape args for last layers of the model.
Instead of the current model definition in unet.py (notice the input_shape arg for the Keras Reshape):



reshape = Reshape((self.img_rows * self.img_cols, 12), input_shape=(self.img_rows, self.img_cols, 12))(conv9)

print("reshape shape:", reshape.shape)



# permute = Permute((2, 1))(reshape)

# print("permute shape:", permute.shape)



activation = Activation('softmax')(reshape)


After 10 epochs I got this result:

[Ajeosi]

Thank you very much!!

[hsinshihlun]

I have trained about 150 epochs, I defined a test visualize which is following below, but the results seems something wrong, colors are same as authors, @Ajeosi @shabtayor how do you visualize your results?
Sky = [128,128,128]
Building = [128,0,0]
Pole = [192,192,128]
Road = [128,64,128]
Pavement = [60,40,222]
Tree = [128,128,0]
SignSymbol = [192,128,128]
Fence = [64,64,128]
Car = [64,0,128]
Pedestrian = [64,64,0]
Bicyclist = [0,128,192]
Unlabelled = [0,0,0]

COLOR_DICT = np.array([Sky, Building, Pole, Road, Pavement,Tree, SignSymbol, Fence, Car, Pedestrian, Bicyclist, Unlabelled])

def test_multi_images(test_path):
dirs = os.listdir(test_path)
for filename in dirs:
print(frame)
img_path = os.path.join(test_path,filename)
img = cv2.imread(img_path)
img_color = img.copy()
dims = img.shape
img = img / 255
img = np.reshape(img,(1,360,480,3))
y_prob = model.predict(img)
y_classes = y_prob.argmax(axis=-1)
output=np.reshape(y_classes,(360,480))
for j in range(dims[0]): #height
for i in range(dims[1]): #width
img_color[j, i] = COLOR_DICT[output[j, i]]
Epoch 147/220
323/323 [==============================] - 54s 167ms/step - loss: 0.0371 - acc: 0.9854