search

matterport / Mask_RCNN

Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow
Other
24.73k stars 11.71k forks source link

Implementation of Mask RCNN in TensorFlow 2.x #2246

Open ulhaqi12 opened 4 years ago

ulhaqi12 commented 4 years ago

Hi, Developer community, I hope you all are in good health. I am using Mask RCNN to build an application and having problems with a version mismatch. because some other things I found online are in TensorFlow 2.x while Mask RCNN is TensorFlow 1.x. It would be great if you'll consider it to provide it in TensorFlow 2.x also. Thanks

saqib1707 commented 4 years ago

Hello @ulhaqi12, You can try upgrading your TF-1.x project to TF-2.x using this official tool provided by TensorFlow here. I did the conversion using this and it works perfectly fine for me in TF-2.x

innat commented 4 years ago

@saqib1707 sorry, but it would not work. @ulhaqi12 best way to downgrade your tensorflow version, (currently) tf. 2.x is not supported.

saqib1707 commented 4 years ago

@innat thanks for your reply. As I mentioned above, it did work for me. Your suggestion also seems good.

ulhaqi12 commented 4 years ago

Thanks for your suggestions, I have successfully tested Mask RCNN keras model on PC. but I am not able to convert it in tflite. There is some custom layers used in this model. I don't know how to deal with that.

burhr2 commented 4 years ago

Hi kindly see this issue https://github.com/matterport/Mask_RCNN/issues/2312 there is a pointer to the same implementation that supports TensorFlow 2=>

alexander-pv commented 3 years ago

Hello, maybe it is outdated but I updated Mask-RCNN for tensorflow v2 and added some more backbones for tests including mobilenet and efficientnet architectures. For now coco weights are not included. You can check this out: link

innat commented 3 years ago

@waleedka This MaskRCNN implementation is vastly used among the tf. keras practitioners still now. But unfortunately, it's been not updated for a long time. It should be updated for tensorflo.keras 2. x. Could you please look into this issue? Is PR welcomed? IMO, the limitation it has now are the a) backbones and the b) advance augmentation i.e cutmix, mixup, etc, and official support for the core API's interface update.

@alexander-pv
That's a really good job. I think it would be better to keep everything in one place. Have you considered PR on that? If @waleedka approves the contribution, please send PR.