Mask R-CNN overview in Computer Vision - Model Pipeline Trace

Model Pipeline - Mask R-CNN overview

Mask R-CNN is a model that finds objects in pictures, draws boxes around them, and colors each object precisely. It does this by first spotting objects, then figuring out their exact shape.

Data Flow - 6 Stages

1Input Image

1 image x 800 x 800 x 3 channels→Original color image loaded→1 image x 800 x 800 x 3 channels

A photo of a dog and a cat

↓

2Backbone CNN

1 image x 800 x 800 x 3→Extracts features like edges and textures→1 image x 50 x 50 x 256 feature maps

Feature maps highlighting shapes of dog and cat

↓

3Region Proposal Network (RPN)

1 image x 50 x 50 x 256→Suggests boxes where objects might be→1000 boxes x 4 coordinates + scores

Boxes around dog ears, cat face, etc.

↓

4RoI Align

1000 boxes x feature maps→Extracts fixed-size features for each box→1000 regions x 14 x 14 x 256

Features for each proposed box

↓

5Bounding Box & Classifier Head

1000 regions x 14 x 14 x 256→Predicts object class and refines box→1000 boxes with class labels and refined coordinates

Box labeled 'dog' with improved position

↓

6Mask Head

1000 regions x 14 x 14 x 256→Predicts pixel-level mask for each object→1000 masks x 28 x 28 pixels

Mask showing exact dog shape inside box

Training Trace - Epoch by Epoch

Loss: 1.2 |****     
Loss: 0.9 |******   
Loss: 0.7 |******** 
Loss: 0.55|*********
Loss: 0.45|**********

Epoch	Loss ↓	Accuracy ↑	Observation
1	1.2	0.45	Model starts learning basic object shapes
2	0.9	0.60	Better detection and mask prediction
3	0.7	0.72	Improved box accuracy and mask quality
4	0.55	0.80	Model refines object boundaries well
5	0.45	0.85	Strong detection and precise masks

Prediction Trace - 6 Layers

Layer 1: Input Image

Layer 2: Backbone CNN

Layer 3: Region Proposal Network

Layer 4: RoI Align

Layer 5: Bounding Box & Classifier Head

Layer 6: Mask Head

Model Quiz - 3 Questions

What does the Region Proposal Network (RPN) do in Mask R-CNN?

A Colors the objects in the image

B Extracts edges and textures

C Suggests boxes where objects might be

D Predicts the object class

Key Insight

Mask R-CNN combines object detection and segmentation by first finding object boxes and then predicting detailed masks, allowing precise understanding of objects in images.