Many of the architectures that do semantic segmentation like SegNet, DilatedNet (Yu and Koltun), DeepLab, etc. do not work on high resolution images. For such benchmarks like Cityscapes, what is a standard/practical approach for such methods to perform on the benchmark?

I've tried to look into the paper, but I couldn't find such details. There's an article mentioning that they output at 1/8 of input images than do interpolation (usually 2, 4 or 8 times) from their results, but the article does not specify which upsampling techniques are the most reasonable one.


You can always train your network with higher resolution images. There is nothing preventing you to do that if you don't have any restriction about inference time.

Also, the paper is actually mentioning how to upsample. Check for the phrase "Upsample using deconvolutional layers".

The most common ways to upsample are either using the deconvolutional layer or just by simple resize methods (like image resize).

| improve this answer | |
  • $\begingroup$ It kinda fascinates me to know that people just employ simple bilinear-interpolation image resize to scale up and down. Do you have experiment how much artifact that would introduce, say image and annotation might be mis-align due to the scaling process? $\endgroup$ – AugLe Feb 28 '18 at 2:26
  • $\begingroup$ Nope. As soon as you scale as floating point, it should be fine. $\endgroup$ – Deniz Beker Feb 28 '18 at 9:56

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.