roi をencodeしたものをfeature map に投影する際, shapeが合わないので工夫する必要がある → ROI pooling と ROI Align (Mask RCNNはコッチ) mask-branchでmaskを生成 各画素ごとにクラス確率を計算 ROI pooling ROI Align bilinear補完を行う ...
headはlossを小さく, tailはlossを大きくしたい 重み $w_i $を使ってlossを設計する (二値の場合) $L_{EQL}=-\sum_{j=1}^{C}w_{j}log(\hat{p_{j}}),$ $w_{j}=1-E(r)T_{\lambda}(f_{j})(1-y_{j})$ In this equation, E(r) outputs 1 when r is a foreground region proposal and 0 when it belongs to background. And fj is the frequency of category j in the dataset, which is computed by the image number of the class j over the image number of the entire dataset. And Tλ(x) is a threshold function which outputs 1 when x < λ and 0 otherwise. λ is utilized to distinguish tail categories from all other categories and Tail Ratio (T R) is used as the criterion to set the value of it TRを元に $\lambda$ を ...
The main differences from [4, 12 are a batch-size of 1024 instead of 4096, the use of global average-pooling (GAP) instead of a class token [2, 11 , fixed 2D sin-cos position embeddings [2, and the introduction of a small amount of RandAugment [3 and Mixup [21 (level 10 and probability 0.2 respectively, which is less than [12). These small changes lead to significantly better performance than that originally reported in [4.
https://arxiv.org/pdf/2205.01580.pdf ...