foamliu/Deep-Image-Matting-PyTorch

my own datasets are all full human body images

Closed this issue · 5 comments

lfxx commented

Hi,thanks for your excellent work.Now i prepare my own datasets.This datasets are consists of thounds of high resolution image(average 4000*4000).They are all full human body images.When i
process these images,i meet a questions:
When i crop the trimap(generated from alpha),often crop some places which are not include hair.Such as foot,leg.Is it ok to input these images into model?@foamliu

lfxx commented

Hi,thanks for your excellent work.Now i prepare my own datasets.This datasets are consists of thounds of high resolution image(average 4000*4000).They are all full human body images.When i
process these images,i meet a questions:
When i crop the trimap(generated from alpha),often crop some places which are not include hair.Such as foot,leg.Is it ok to input these images into model?@foamliu

i tried to train the model using those samples,find that model hard to converage.In my opinion,the images inputted to model must include hair.Am i right?

polly-morphism commented

Hi,thanks for your excellent work.Now i prepare my own datasets.This datasets are consists of thounds of high resolution image(average 4000*4000).They are all full human body images.When i
process these images,i meet a questions:
When i crop the trimap(generated from alpha),often crop some places which are not include hair.Such as foot,leg.Is it ok to input these images into model?@foamliu

i tried to train the model using those samples,find that model hard to converage.In my opinion,the images inputted to model must include hair.Am i right?

Hi. I have a similar dataset (2k full body images with masks 2.5k*3k pixels). Were you able to successfully train the model? I don't have hair in my dataset at all. How long did it take to train the model? What GPU did you use? I will be so grateful for any information)

xiankgx commented

These CNN models are usually not trained on full sized images. For 2k images, and around 50 epochs, training will probably only take a few hours on a GPU.

xiankgx commented

If you don't mind, can you share your dataset?

foamliu commented

The model is highly dependent on trimaps.