Dai et al. Propose Convolutional Feature Masking (CFM) for instance segmentation. A CFM layer tackles the following two problem: instead of applying masks or bounding boxes on the image content, thereby creating artifacts and requiring multiple forward passes, the convolutional feature maps are computed first and then masked within the CFM layer. In their work, the object proposal masks are obtained from a state-of-the art object proposal generator such as Selective Search (SS)  or Multiscale Combinatorial Grouping (MCG) . This difference is illustrated in Figure 1 and Figure 2 shows the idea of the CFM layer.
Figure 1: Illustration of the proposed approach in comparison to state-of-the-art approaches. The Convolutional Feature Maskign (CFM) layer allows to first compute all feature maps and then use proposals to mask the features.
They propose two different ways to integrate the CFM layer with an SPP layer  as shown in Figure 3. In experiments, they show that both approaches perform similarly and therefore propose to use Design B (right in Figure 3) as it requires less computation.