Learning to Segment Every Thing論文導讀

kaiming較新的工作，剛剛有時間簡單看一遍。

主要的思想是解決目前instance segmentation的標註數據太少且標註成本太高, 導致目前instance的分割類別較少，而不計一切地大量標註也不是好的解決辦法。

於是在mask rcnn的基礎上提出了mask X rcnn. 主要思想是遷移學習的思想，將mask rcnn detection的branch和mask branch相連接, 將detection 的特徵引入到mask branch中。

Genome數據集有大量的detection bbox標註, 而只有coco DataSets有稀有的instance 標註, 所以做法是利用Genome和coco的detection bbox groundtruth. 利用coco DataSets 的instance groundtruth。

因為訓練方式的話，mixture of strongly annotated examples (those with masks) and more weakly annotated examples (those with only boxes),所以文章稱之為"partially supervised".

Let C be the set of object categories.

Assume that C = A ∪ B where examples from the categories in A have masks, while those in B have only bounding boxes.

在Mask-Rcnn里, bbox和mask branch最後的layer都是category-specific 的參數, 用來執行bounding box classification and instance mask prediction。

Instead of learning the category-specific bounding box parameters and mask pa- rameters independently, we propose to predict a category』s mask parameters from its bounding box parameters using a generic weight transfer function that can be jointly trained as part of the whole model.

訓練的話還是分階段訓, 先train detection branch，然後再end-to-end train.

推薦閱讀：