search

DrSleep / tensorflow-deeplab-resnet

DeepLab-ResNet rebuilt in TensorFlow
MIT License
1.25k stars 431 forks source link

coco dataset pretraining #48

Closed gaopeng-eugene closed 7 years ago

gaopeng-eugene commented 7 years ago

Recently, I am trying to add coco dataset into this framework which is a common trick for many state-of-the-art semantic segmentation method. I will try to add the code as soon as possible.

arslan-chaudhry commented 7 years ago

If I am not mistaken init.caffemodel which is converted to deeplab_resnet_init.ckpt is already pre-trained on MS-COCO for semantic segmentation task. @DrSleep can you please confirm that? Also, do you have resenet init model trained for classification task converted to tf .ckpt format somewhere? If not, do you know if Deeplab folks have released any such init model? I'm trying to replicate the results for weakly supervised setting and I don't want pre-training on fully supervised coco dataset.

DrSleep commented 7 years ago

Yes, it is pre-trained on COCO for sem. segmenation.

I don't have the classification net in ckpt, unfortunately. DeepLab folks have been using the files provided by K.He pre-trained on ImageNet. I think you may use them as well with slight changes in prototxt if necessary. On 16 Mar 2017 6:03 pm, "Arslan Chaudhry" notifications@github.com wrote:

If I am not mistaken init.caffemodel which is converted to deeplab_resnet_init.ckpt is already pre-trained on MS-COCO for semantic segmentation task. @DrSleep https://github.com/DrSleep can you please confirm that. Also, do you have resenet init model trained for classification task converted to tf .ckpt format somewhere. If not, do you know if Deeplab folks have released any such init model? I'm trying to replicate the results for weakly supervised setting and I don't want pre-training on fully supervised coco dataset.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/DrSleep/tensorflow-deeplab-resnet/issues/48#issuecomment-287085565, or mute the thread https://github.com/notifications/unsubscribe-auth/AHemmBKG-F888d01SXwkVrpfocO91Fduks5rmU8sgaJpZM4MTalx .

bhack commented 7 years ago

What about sceneparsing?

gaopeng-eugene commented 7 years ago

I think you are right. Joint training COCO and PASCAL VOC give almost the same performance.

arslan-chaudhry commented 7 years ago

@gaopeng-eugene if you have the segmentation masks and class categories of coco dataset extracted somewhere from json files, could you please share those?