Mask R-CNN

The paper extends Faster R-CNN with a parallel mask prediction branch and a better feature alignment operator for instance segmentation.

Metadata

Field	Value
Paper	Mask R-CNN
Authors	Kaiming He, Georgia Gkioxari, Piotr Dollár, Ross Girshick
Year	2017
Venue	ICCV 2017
arXiv	1703.06870
Status	seed note started

One-Line Takeaway

Mask R-CNN keeps the two-stage detection structure of Faster R-CNN and adds an instance-specific dense mask head, making detection, box regression, and mask prediction share one region-based architecture.

Question

Object detection predicts boxes:

$${\hat{Y}}_{\det }=\{({\hat{c}}_i,{\hat{b}}_i,{\hat{s}}_i){\}}_{i=1}^N.$$

Instance segmentation predicts boxes plus a separate mask for each object:

$${\hat{Y}}_{\mathrm{inst}}=\{({\hat{c}}_i,{\hat{b}}_i,{\hat{m}}_i,{\hat{s}}_i){\}}_{i=1}^N,$$

where:

• ${\hat{c}}_i$ is the class;
• ${\hat{b}}_i$ is the bounding box;
• ${\hat{m}}_i\in [0,1{]}^{H_m\times W_m}$ is the instance mask;
• ${\hat{s}}_i$ is the confidence score.

The paper asks:

Can a strong two-stage detector be extended to pixel-level instance masks without making the architecture complicated?

Architecture Contract

Component	Role
CNN backbone	extracts image features
Region Proposal Network	proposes object-like regions
RoIAlign	extracts spatially aligned region features
classification head	predicts object category
box head	refines bounding box coordinates
mask head	predicts a class-specific binary mask for each RoI

The architectural change from Faster R-CNN:

$$\text{region detection head}\to \text{region detection head}+\text{parallel mask head}.$$

RoIAlign

Faster R-CNN uses RoI pooling to convert variable-sized regions into fixed-size tensors. Quantization in this step can misalign features and pixels, which hurts mask quality.

Mask R-CNN replaces it with RoIAlign:

$$z_i=\mathrm{RoIAlign}(F,r_i),$$

where $F$ is the feature map and $r_i$ is a proposed region.

Conceptually:

1. avoid rounding region coordinates to coarse feature bins;
2. sample feature values at fractional coordinates;
3. use bilinear interpolation;
4. preserve spatial alignment for the mask head.

For a fractional coordinate $(x,y)$, bilinear interpolation can be written as:

$$F(x,y)=\sum _{u,v}{w}_{uv}(x,y)F(u,v),$$

where neighboring grid points $(u,v)$ contribute according to interpolation weights $w_{uv}$.

This is a small architectural operator, but it matters because masks are evaluated at pixel-level boundaries.

Mask Head

For each region feature $z_i$, Mask R-CNN predicts:

$${\hat{m}}_i=f_{\mathrm{mask}}(z_i)\in [0,1{]}^{K\times H_m\times W_m},$$

where $K$ is the number of classes.

The selected mask is usually the channel corresponding to the predicted class:

$${\hat{m}}_{i,{\hat{c}}_i}\in [0,1{]}^{H_m\times W_m}.$$

The mask branch is fully convolutional, so it preserves spatial structure inside the RoI:

$$z_i\to \mathrm{Conv}\to \mathrm{Conv}\to {\hat{m}}_i.$$

Multi-Task Objective

Mask R-CNN uses the Faster R-CNN detection losses plus a mask loss:

$$\mathcal{L}={\mathcal{L}}_{\mathrm{cls}}+{\mathcal{L}}_{\mathrm{box}}+{\mathcal{L}}_{\mathrm{mask}}.$$

For the target class $k^{\text{*}}$, the mask loss is binary cross entropy over the mask grid:

$${\mathcal{L}}_{\mathrm{mask}}=-\frac{1}{H_m{W}_m}\sum _{h,w}\left[m_{h,w}^{\text{*}}\log {\hat{m}}_{k^{\text{*}},h,w}+(1-m_{h,w}^{\text{*}})\log (1-{\hat{m}}_{k^{\text{*}},h,w})\right].$$

This keeps mask prediction class-specific while avoiding competition between classes at every pixel.

Why It Matters

Mask R-CNN is a canonical example of extending an architecture by adding a task head while preserving the core representation pipeline:

$$\text{shared image features}\to \text{region features}\to \{\text{class},\text{box},\text{mask}\}.$$

The important lesson is not only that masks improve segmentation. It is that a clean architecture can support several structured outputs if the intermediate representation has the right granularity.

Relation To Later Models

Model family	Object representation	Output style
Faster R-CNN	proposed RoIs	class and box
Mask R-CNN	proposed RoIs	class, box, and mask
DETR	learned object queries	class and box set
modern segmenters	prompts, masks, or query slots	dense masks or mask sets

Mask R-CNN is still useful to read before modern promptable segmentation systems because it makes the difference between detection boxes, aligned RoI features, and per-instance masks explicit.

What To Watch

• Mask quality depends heavily on feature alignment.
• The model inherits two-stage detector dependencies: proposals, anchor choices, and NMS.
• The mask head predicts masks per region, not a single global semantic map.
• Box AP and mask AP can move differently; one does not fully explain the other.
• The method is architecture-general: stronger backbones and feature pyramids can be plugged in.

Related

• Object detection
• Segmentation
• Localization
• CNN
• Feature Pyramid Networks
• Faster R-CNN
• DETR