Anyone came across that the upper left corner of the image appears "purple".

[xiaozhi2015]

[xiaozhi2015]

And it appears in almost every result image.

[KupynOrest]

There is a bug in the inference, will be updated soon

[xiaozhi2015]

Thx for your quick reply, looking forward to the update...

[vgg4resnet]

yes，same issue here,has this been solved? I test my images ，purple filled in full image， not only the corner of images， is it this caused by np.transpose function?

[vgg4resnet]

so is that caused by the network is not converged？my predict result images were like this

[vgg4resnet]

my training images are gray ，I convert them to BGR format,I also remove the hsv,rgb,transform part by changing the config.yaml

the result keep the same ,anywhere maybe lead the problem?By the way I also remove sharpen part，we are bluring the images,why do we need that?I am confused about that

[t-martyniuk]

please check the updated links to models (latest changes to readme).
should work fine now

[xiaozhi2015]

Much thx.

[vlee-harmonicinc]

There is still some purple artifact on images with new pre-trained model.
For example:

It seems that relatively dark region/texture prone to have this artifact. When generating video, it trend to have artifact on same object (but not every frames)
Do anyone have idea why this issue happen?

[Sandbox3aster]

@htleeab Which image in GoPro did you test on and which model did you use? Let me see if i can reproduce that

[vlee-harmonicinc]

@Sandbox3aster
Those image are my own test assets, not GoPro dataset. I used the pretrained InceptionResNet-v2 (md5sum: 96f747f38a0119669265cbb5fc7b3c5c)






Thanks.

[Johnreidsilver]

Same problem here, using cpu and same model
96f747f38a0119669265cbb5fc7b3c5c best_fpn.h5

Ubuntu 18.04.02, python 3.6.9
torch==1.5.0
torchsummary==1.5.1
torchvision==0.6.0
numpy==1.17.4
opencv-python-headless==4.2.0.34
joblib==0.14.0
albumentations==0.4.5
scikit-image==0.17.2
tqdm==4.46.0
tensorboard==2.1.1
fire==0.3.1

Perhaps the newer torch version is acting up?