MGCAM icon indicating copy to clipboard operation
MGCAM copied to clipboard

How do you get a foreground image without background

Open gakkitw opened this issue 6 years ago • 12 comments

How do you use FCN to get a foreground image without background? I am a newcomer to this, can you give me some advice? I only want to get a pedestrian image without background. What should I do? Thank you.

gakkitw avatar Mar 11 '19 19:03 gakkitw

We get the foreground image with a well-trained FCN model. We are sorry that we could not share this segmentation model. You can refer to the paper of "Early Hierarchical Contexts Learned by Convolutional Networks for Image Segmentation." and "1000fps human segmentation with deep convolutional neural networks" to find some details.

developfeng avatar Mar 13 '19 06:03 developfeng

Hi, I have read that 2 papers and download the baidu segmentation dataset. I trained it with a FCN model, however, the test on the ReID dataset are not good enough. So, have you used some tricks such as preprocess the baidu dataset? Can you share some details? Thank you so much! @developfeng

hjwang-824 avatar Apr 07 '19 10:04 hjwang-824

Hi, there are some differences with above two paper. 1) We re-design and modify the FCN based model motivatied by above two papers. However, the model used in our CVPR paper is not exactly the one in "1000fps human segmentation with deep convolutional neural networks". We are sorry that we can not share the segmentation model now. 2) Similar with the way in the paper "1000fps human segmentation with deep convolutional neural networks", much more extra data is added for better training.

developfeng avatar Apr 09 '19 04:04 developfeng

Thank you,I will try again.

hjwang-824 avatar Apr 09 '19 17:04 hjwang-824

And can you share your extra data if it’s OK to you?

hjwang-824 avatar Apr 09 '19 18:04 hjwang-824

Sorry, we can't share the segmentation data. I think one possible solution is that, you can try to select some released high-quality masks and its RGB images of Market-1501 and MARS to train your model.

developfeng avatar Apr 10 '19 03:04 developfeng

OK,thank you.

hjwang-824 avatar Apr 10 '19 03:04 hjwang-824

@Ko8e-24 ... I Do you mean how to crave out the background and keep the foreground( The person ) ? If so , I guess the mask shared in this repository by @developfeng and the original RGB image in can do that for you. foreground = RGB_Image x Mask_of_RGB_Image. I tried and here is some sample I got : -

original_forehround_mask

@developfeng Thank you for sharing the mask for Market-1501 and cuhk03. I understand you didn't start sharing the segementation code you used to get your masks, but I would like to know if you have mask data for DukeMTMC dataset

bmiftah avatar Jun 13 '20 13:06 bmiftah

hello,could you share the mask data for DukeMTMC dataset? Thank you very mach.

HWNHJJ avatar Nov 17 '20 11:11 HWNHJJ

@HWNHJJ I am also looking for these dataset ... hope @developfeng could share us sooner .

bmiftah avatar Nov 17 '20 13:11 bmiftah

@developfeng, is the code for the paper "1000fps human segmentation with deep convolutional neural networks" available somewhere?

SamSamhuns avatar Dec 10 '20 08:12 SamSamhuns

@Ko8e-24 ... I Do you mean how to crave out the background and keep the foreground( The person ) ? If so , I guess the mask shared in this repository by @developfeng and the original RGB image in can do that for you. foreground = RGB_Image x Mask_of_RGB_Image. I tried and here is some sample I got : -

original_forehround_mask

@developfeng Thank you for sharing the mask for Market-1501 and cuhk03. I understand you didn't start sharing the segementation code you used to get your masks, but I would like to know if you have mask data for DukeMTMC dataset

Thank you for sharing method to get a foreground, I do as you told ,but why I get this?And ,why your mask is yellow-purple, since author gave us black-white mask? fg

BonBon-Boom avatar Mar 14 '23 13:03 BonBon-Boom