This is the Mask R-CNN paper. It extends Faster R-CNN to produce pixel masks, i.e., for semantic segmentation. It extends Faster R-CNN by adding a branch for predicting segmentation masks on each Region of Interest (RoI). However, this have to support with RoIAlign instead of RoIPooling, to avoid error introduced by quantization.

The key contribution of design is to point out the importance of decoupling mask and class prediction, i.e., inference of the binary mask should not depend on classification result. The Mask R-CNN model runs these two tasks in parallel.

## Architecture

In Faster R-CNN, features are extracted from feature map based on a region proposal (RoI). The feature vector is obtained from RoIPool layer, which the feature vector is feed into classification and bounding-box regression.

In Mask R-CNN, the RPN stage is same as Faster R-CNN, but afterwards, the classification, box offset, and binary mask prediction are run in parallel for each RoI. The classification does not depend on mask prediction as in other model, but performed in parallel.

The classification and bounding-box regression is same as that in Faster R-CNN, but the mask prediction is to produce a $$Km^2$$-dimensional output for each RoI, for $$K$$ binary masks of resolution $$m\times m$$. It is implemented as per-pixel sigmoid using FCN.

Mask prediction requires per-pixel spatial correspondence, hence we need the feature maps to aligned to pixels. This is done by RoIAlign: It is to produce a small feature map (e.g., 7×7) from an arbitrary size RoI. The algorithm is as follows:

1. Given RoI, subdivide its spatial dimension $$x$$ into bins of size $$x/m$$ (a floating point)
2. In each bin, calculate the value at, say, 4 sampling points. The 4 sampling points are the center intersections by drawing 3×3 grids on the bin. The value is computed by bilinear interpolation with neighboring pixels.
3. The value assigned to this bin is aggregated (e.g., max or average) from the sampling points

In training this model, the loss function is $$L=L_{\text{cls}}+L_{\text{box}}+L_{\text{mask}}$$, which the first two are the classification and bounding-box regression loss as in Faster R-CNN, and the last one is the mask prediction loss defined as the average binary cross-entropy loss of each pixel.

## Bibliographic data

@inproceedings{
}