Image resize for training

[QuickLearner171998]

Hi, I did not find an option to resize images for training.

Also while inferencing if I resize the images to 384 by changing these lines

text-detection-fots.pytorch/test.py

Line 20 in 04b969c

scale_x = 2240 / image.shape[1] # 2240 # 1280

nothing is predicted using the pretrained model.

[Wovchena]

Training images are upscaled here:

text-detection-fots.pytorch/datasets.py

Line 32 in 04b969c

upscaled = cv2.resize(im, None, fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC)

I think the model doesn't work for size of 384 because it was trained for a larger scale. You may also try visualizing raw net outputs. May be it will give you some understanding why it happens.

[QuickLearner171998]

Hey I tried training on images with size 384x384 , train loss curve was also fine.
But if I resize test images(as mentioned above) the results are incorrect and when I keep the code scales as it is I get decent results. Any idea why is this happening?

I notice here:

text-detection-fots.pytorch/datasets.py

Line 25 in 04b969c

IN_SIDE = 640

IN_SIDE = 640 does that mean the input size in training is 640?

[QuickLearner171998]

Training images are upscaled here:

text-detection-fots.pytorch/datasets.py

Line 32 in 04b969c

upscaled = cv2.resize(im, None, fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC)

I think the model doesn't work for size of 384 because it was trained for a larger scale. You may also try visualizing raw net outputs. May be it will give you some understanding why it happens.

Hi I am trying to visualize the gt masks...I changed this line:

text-detection-fots.pytorch/train.py

Line 211 in 04b969c

if 0:

to if 1
Only black images are shown and now there is div by zero error while calculating dist.
Also, can you explain how can I visualize the masks in between training?