How can i use the model to predict for bigger image like 720p?

need i to train using the same size image,like 720p?

I have resolved it.
By now I think you should fine-tuning the models to make it fit for big size image,but the m_iou decrease.

Hi George ... Could you pls elaborate on how to resolved it? I am stuck here!

我已经解决了。
到目前为止，我认为你应该对模型进行微调以使其适合大尺寸图像，但是m_iou会减少。
hi.did u like to share me how to solve this problem?

嗨乔治......你能详细说明如何解决它吗？我被困在这里！

Did u solve it? I also has the problem

No, i couldn't

I met this problem after train deeplab model.
This is caused by, deferent image size of train and test.
For example, train with (513,513), in input_preprocess.py preprocess_image_and_label function.
if is_training and label is not None

will crop the image.
but in test phase, is_training is False.
So the image size will not be changed.
So my temple solution is that, add else branch:

else:
  rr = tf.minimum(tf.cast(crop_height,tf.float32)/tf.cast(image_height,tf.float32),\
           tf.cast(crop_width,tf.float32)/tf.cast(image_width,tf.float32))
   newh = tf.cast(tf.cast(image_height, tf.float32)*rr, tf.float32)
   neww = tf.cast((tf.cast(image_width, tf.float32)*rr), tf.float32)
   processed_image = tf.image.resize_images(
       processed_image, (newh, neww), method=tf.image.ResizeMethod.BILINEAR, align_corners=True)
   processed_image = preprocess_utils.pad_to_bounding_box(
                       processed_image, 0, 0, crop_height, crop_width, mean_pixel)

*In detail, if you choose xception, in xception.py:
inputs = fixed_padding(inputs, kernel_size, rate)

I met this problem after train deeplab model.
This is caused by, deferent image size of train and test.
For example, train with (513,513), in input_preprocess.py preprocess_image_and_label function.
if is_training and label is not None

will crop the image.
but in test phase, is_training is False.
So the image size will not be changed.
So my temple solution is that, add else branch:

else:
  rr = tf.minimum(tf.cast(crop_height,tf.float32)/tf.cast(image_height,tf.float32),\
           tf.cast(crop_width,tf.float32)/tf.cast(image_width,tf.float32))
   newh = tf.cast(tf.cast(image_height, tf.float32)*rr, tf.float32)
   neww = tf.cast((tf.cast(image_width, tf.float32)*rr), tf.float32)
   processed_image = tf.image.resize_images(
       processed_image, (newh, neww), method=tf.image.ResizeMethod.BILINEAR, align_corners=True)
   processed_image = preprocess_utils.pad_to_bounding_box(
                       processed_image, 0, 0, crop_height, crop_width, mean_pixel)

*In detail, if you choose xception, in xception.py:
inputs = fixed_padding(inputs, kernel_size, rate)

it works for the eval.py and vis.py of deeplab!
but one mistake: the cast of neww and newh should be tf.int32
wrong:

newh = tf.cast(tf.cast(image_height, tf.float32)*rr, tf.float32) 
neww = tf.cast((tf.cast(image_width, tf.float32)*rr), tf.float32)

right:

newh = tf.cast(tf.cast(image_height, tf.float32)*rr, tf.int32) 
neww = tf.cast((tf.cast(image_width, tf.float32)*rr), tf.int32)

and the visualize result is crop region of original image which may not be the purpose of the auther. here is a reference about the explaination of the auther:
https://github.com/tensorflow/models/issues/3939

@northeastsquare , @muxizju
Big thanks guys, now it works. After hours of struggling your else branch did the job :)

I met this problem after train deeplab model.
This is caused by, deferent image size of train and test.
For example, train with (513,513), in input_preprocess.py preprocess_image_and_label function.
if is_training and label is not None

will crop the image.
but in test phase, is_training is False.
So the image size will not be changed.
So my temple solution is that, add else branch:

else:
  rr = tf.minimum(tf.cast(crop_height,tf.float32)/tf.cast(image_height,tf.float32),\
           tf.cast(crop_width,tf.float32)/tf.cast(image_width,tf.float32))
   newh = tf.cast(tf.cast(image_height, tf.float32)*rr, tf.float32)
   neww = tf.cast((tf.cast(image_width, tf.float32)*rr), tf.float32)
   processed_image = tf.image.resize_images(
       processed_image, (newh, neww), method=tf.image.ResizeMethod.BILINEAR, align_corners=True)
   processed_image = preprocess_utils.pad_to_bounding_box(
                       processed_image, 0, 0, crop_height, crop_width, mean_pixel)

*In detail, if you choose xception, in xception.py:
inputs = fixed_padding(inputs, kernel_size, rate)

what does In detail mean? need to change something ?

@mohhao Did you figure out what to do with the xception.py file?

@northeastsquare Can you explain the changes to be maid to xception.py file?

I think instead of doing this hack you should just set the min_size and max_size flags both to the crop size. And keep the default flag value for preserving the aspect ratio.

This will result in scaling the larger dimension to the crop size, then padding to fill the rest.

*I mean set the flags

--min_resize_value=513
--max_resize_value=513