Hello, Thank you for sharing your great dataset. When I downloaded dataset all Mask images were black here no masked imaged. Where I can find the masked images?
Following are the results of my execution using your code (no modification in the hyperparameter and model):
model at 40 epoch,
(train) bin_accuracy: 0.9577
val_bin_accuracy: 0.9514
test_accuracy : 0.5724729895591736
model at 37 epoch (best model based on val set),
(train) bin_accuracy: 0.9568
val_bin_accuracy: 0.9820
test_accuracy : 0.6100730895996094
confusion matrix:
[[ 814. 2666.]
[ 694. 4443.]]
train and validation accuracy is actually similar enough, but the test accuracy is quite far from the paper.
When I implemented your code and re-run it, I got different IoU values from the ones that you reported in the paper.
There is huge difference between these values i.e., you reported 78.17 IoU while I achieved 49.72 value of IoU.
The mask files given in the dataset are black images as corresponding to the input image data. Can you please upload the actual binary masks related to the input aerial images?
I download the segmentation dataset,but I do not know the specific splitting file for train,val,test set. Please provide the relevant splitting file. Thank you very much for your help.
Hi,
since I'm working on NVIDIA Jetson Nano developer kit, and it cannot perform the training of the model due to RAM limitations, is there any chance you can provide the trained model so I can run just the segmentation script? I tried so, but it isn't possible without the training process.
Thanks in advance.