dulucas / refinement_net

Train a class-agnostic segmentation network which takes an cropped image as input

Box2Seg network

This is an implementation of Box2Seg network proposed in paper(Section 3.3). Given an image with a bounding box, the network learns to predict a corresponding binary mask for the object inside the bounding box.

Architecture

ResNet101-FPN

Acknowledgement

Codes based on TorchSeg and Detectron2

Contact

Feel free to contact me if you have any questions : [e-mail]