search

matterport / Mask_RCNN

Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow
Other
24.66k stars 11.7k forks source link

Can this implementation achieve the state-of-art performance? #1412

Open wuyangzhang opened 5 years ago

wuyangzhang commented 5 years ago

Hi Guys,

I tested the implementation on the COCO dataset with the pre-trained weights, while I cannot get the state of art performance shown at http://cocodataset.org/#detection-leaderboard. At this point, I am trying to figure out what is going wrong.. Is there any one who can achieve the expected performance with the implementation?

ppwwyyxx commented 5 years ago

It cannot. See #1.

AndreaPi commented 5 years ago

@ppwwyyxx interesting. I guess you're using another implementation, then. May I ask which one?

ppwwyyxx commented 5 years ago

I use my own implementation in TensorFlow, and maskrcnn-benchmark in PyTorch as well.

Not saying that they reach best performance on COCO leaderboard, but at least they are able to reproduce performance in the official implementation in Detectron.

AndreaPi commented 5 years ago

Nice! Is it as easy to modify your implementation (or maskrcnn-benchmark) to train on a custom dataset, as it is for this one?

ppwwyyxx commented 5 years ago

They have their own documentation 1 2 about adding datasets.