@krishna7611 wrote:
I am currently working on an Image segmentation project. The images I am working on are 4k resolution and the number of background pixels are extremely large compared to foreground pixels. Since it is not feasible to train on such large images I am breaking the image into random crops of 256x256. When I do random cropping I create equal number of images with objects in it and just the background so that I balance the foreground and background. I am using Jaccard distance loss. My training and validation loss curves look very good validation and training loss(please see below). I am predicting on the 4k image by breaking the image into 256x256 patch and reassembling it. But, I see a lot of false positives. Any suggestions would be appreciated.
Posts: 1
Participants: 1