Learning-Deep-Learning

Object Detection based on Region Decomposition and Assembly

@YuShen1116

May 2019

tl;dr

This paper propose a region decomposition and assembly(R-DAD) for more accurate object detection. the motivation of decomposition object is to tackle occlusion: if your target object is blocked by something, decompose the object would help the network to learn the features from unblocked(valid) parts.

Overall impression

This paper is also working on the feature extraction part of the network to improve detection accuracy. In the beginning they are still using the CNN network as backbone, and they applied a multi-scale region proposal and Roi pooling to crate a base feature map. After that, they decompose the input object features into four parts(up, bottom, left, right) and use a region assembly block(RAB) to extract and fuse features for those sub-images, the connect the block to the detection layer.

Key ideas

Technical details

Notes