Semantic segmentation for large images

Question

I am working on a limited number of large images, each of which may have 3072*3072pixels. To train the semantic segmentation model using FCN or U-net, I will build a large selection of training sets, each training image 128*128.

At the forecasting stage, I do this in order to cut a large image into small pieces, just like a training kit 128*128, and feed these small pieces into a trained model, to get a predicted mask. Subsequently, I simply stitch these small patches to get a mask for the whole image. Is this the right mechanism to perform semantic segmentation against large images?

Answer 1

Input Image Data: I would not recommend feeding a large image (3072x3072) directly into coffee. A package of small images will fit better into memory, and parallel programming will come into play as well. It will also be possible to increase data.

Output for large image: As for output of large image, you better remake the input FCN size to 3072x3072 during the testing phase. Because FCN layers can accept inputs of any size. Then you get a 3072x3072 segmented image as output.

Answer 2

. , , , . , kaggle.

, , , . , , .

, . . , .

- , , : , , . , .

curio1729, . , , . CNN , , , .