single dataset baseline

provides code for simple binary segmentation on dataset with strict structure

usage

python3 run.py [path]

• run train.py on [path] which produces a model model.pth on a local build folder - using only data from [path/train/]
• run test.py which uses the model on [path/test/]

data structure

[path] should be the path to a dataset

• [path] should contain 2 folders [train] and [test]
• each [train], [test] should contain a sequence of pairs of image/mask
• image names should be {i}_x.png and mask name should be {i}_y.png
• with i a range of integer starting from 0
• all other data are ignored (in particular if there is "0_x.png, 1_x.png, 3_x.png" then "3_x.png" is ignored as the range stop at 1 but if 2_x.png is added then all images are used)
• mask should be image 0 corresponds to background and other value to the class of interest
• you could find in potsdam.zip a post processed version of the famous ISPRS POTSDAM dataset with suitable data structure

feature

training relies on crops from images

testing processes the hole images by tilling WITH OVERLAP

in torchvision code, first crops are extracted, then pushed into a classical pytorch dataloader

in thread code, cropextractor is threaded so 2 different cropextractors can be used on 2 datasets having very different size however this comes with worse running time compared to torchvision dataloader - and - sampling with lower "randomness"

typical output on Potsdam

load data define model train tensor([1.4680], device='cuda:0') tensor([0.6558], device='cuda:0') tensor([0.4999], device='cuda:0') tensor([0.4049], device='cuda:0') tensor([0.3506], device='cuda:0') tensor([0.3031], device='cuda:0') tensor([0.2750], device='cuda:0') tensor([0.2678], device='cuda:0') tensor([0.2412], device='cuda:0') tensor([0.2263], device='cuda:0') 999 perf tensor([81.8246, 92.2427]) tensor([0.2115], device='cuda:0') tensor([0.2054], device='cuda:0') tensor([0.1975], device='cuda:0') tensor([0.1916], device='cuda:0') tensor([0.1727], device='cuda:0') tensor([0.1735], device='cuda:0') tensor([0.1644], device='cuda:0') tensor([0.1611], device='cuda:0') tensor([0.1548], device='cuda:0') tensor([0.1523], device='cuda:0') 1999 perf tensor([93.5353, 97.5884]) training stops after reaching high training accuracy load data load model test perf= tensor([85.9171, 93.3978]) total duration 989.1485934257507

durations on Potsdam are similar for both versions