search

matterport / Mask_RCNN

Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow
Other
24.37k stars 11.65k forks source link

mask RCNN in production #1087

Open abhrau opened 5 years ago

abhrau commented 5 years ago

Hi All, have few questions regarding usage of mask Rcnn for small applications...I am asking these questions because it seems pretty slow and needs lot of memory to process.

1) Is it possible to use mask RCNN on desktops (no GPUs) for small applications using open CV? 2) If yes, what woudl be the configuration of the target system?

Please share your experiences / thoughts on these..

fastlater commented 5 years ago

Check this https://github.com/matterport/Mask_RCNN/wiki

abhrau commented 5 years ago

Hi fastlater, Thank you very much for your response... so it is clear from the post shared by you it is not a memory efficient model.. but are there any models which are accurate and memory efficient (means fast) compared to mask RCNN? (i am still searching but not yet successful ) And one more question.. when i am trying to test with mask RCNN model, some times it is generating masks for back ground shadows also.. so how to generate the masks for the exact objects? Surely appreciate your inputs Thanks Abhrau

fastlater commented 5 years ago

@abhrau remember that accuracy and time will be inverse proportional almost every time, especially in deep learning. I heard about a Fast Mask-RCNN. Look for that and maybe you will find it. I create mask images. In that way, I need to save those images in a folder (a few gbs) but at least, I am sure masks are exactly what it suppose to be. I dont know if it is what you was asking about.

mehditlili commented 5 years ago

OpenCV can do that for you https://github.com/opencv/opencv/wiki/TensorFlow-Object-Detection-API It takes around 4 seconds per inference on 640x480 Images on an I7 CPU

yamuna83 commented 5 years ago

@mehditlili: I am having the same issue what @woodhead007: having. anyone resolved this issue?

Arham-Aalam commented 5 years ago

@abhrau did you found anything better?