Relation Mask R-CNN

This is an implementation of a Relation Mask R-CNN based on [3]. In this repository, we use a special version of Graph Networks implemented inside Mask R-CNN[2], which called GPI (Graph Permutation Invariant[1]), that use the structure of graphs for a detection task on Berkeley Deep Drive dataset.

Introduction

Qualitative Results

References

[1] Roei Herzig, Moshiko Raboh, Gal Chechik, Jonathan Berant, Amir Globerson, Mapping Images to Scene Graphs with Permutation-Invariant Structured Prediction, NIPS, 2018.

[2] Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick, Mask R-CNN, ICCV, 2017.

[3] Abdulla Waleed, Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow, 2017.

Citation

Please cite our git repo if you use this code in your own work:

@misc{herzig_relation_maskrcnn,
  title={Relation Mask R-CNN for object detection based on Graph Permutation Invariant},
  author={Herzig, Roei},
  year={2018},
  publisher={Github},
  journal={GitHub repository},
  howpublished={\url{https://github.com/roeiherz/RelationMaskRCNN}},
}