Recurrent Pixel Embedding for Instance Grouping

We introduce a differentiable, end-to-end trainable framework for solving pixel-level grouping problems such as instance segmentation consisting of two novel components. First, we regress pixels into a hyper-spherical embedding space so that pixels from the same group have high cosine similarity while those from different groups have similarity below a specified margin. We analyze the choice of em

[elu_18]

インスタンスセグメンテーションなどに見られるピクセル単位のグループ問題を解くために、ピクセル毎に高次元球面上への埋め込みを行い、次に球面上でMean shiftクラスタリングをRNNで実現する手法を提案。このモデルは