OpenPose overview in Computer Vision - Model Pipeline Trace

Model Pipeline - OpenPose overview

OpenPose is a system that detects human body parts and their positions in images or videos. It finds key points like elbows, knees, and wrists to understand human poses.

Data Flow - 5 Stages

1Input Image

1 image x 368 height x 368 width x 3 channels→Resize and normalize the input image→1 image x 368 height x 368 width x 3 channels

A photo of a person standing in a room

↓

2Feature Extraction

1 image x 368 x 368 x 3→Pass image through CNN layers to extract features→1 image x 46 x 46 x 128 feature maps

Feature maps highlighting edges and textures of the person

↓

3Part Confidence Maps

1 image x 46 x 46 x 128→Predict heatmaps showing likelihood of each body part at each location→1 image x 46 x 46 x 18 (body parts)

Heatmap with bright spots where wrists and elbows likely are

↓

4Part Affinity Fields

1 image x 46 x 46 x 128→Predict vector fields showing connections between body parts→1 image x 46 x 46 x 36 (connections)

Vector fields pointing from elbow to wrist

↓

5Pose Assembly

Confidence maps and affinity fields→Combine detected parts and connections to form full body poses→List of poses with keypoint coordinates

Coordinates of detected wrists, elbows, knees for each person

Training Trace - Epoch by Epoch

Loss:
2.5 |*****
1.2 |****
0.7 |***
0.4 |**
0.3 |*

Epochs ->

Epoch	Loss ↓	Accuracy ↑	Observation
1	2.5	0.30	Model starts learning basic body part locations
5	1.2	0.55	Confidence maps and affinity fields improve
10	0.7	0.75	Model detects body parts more accurately
15	0.4	0.85	Pose assembly becomes reliable
20	0.3	0.90	Model converges with good pose detection

Prediction Trace - 5 Layers

Layer 1: Input Image

Layer 2: Feature Extraction CNN

Layer 3: Part Confidence Maps Prediction

Layer 4: Part Affinity Fields Prediction

Layer 5: Pose Assembly

Model Quiz - 3 Questions

What does the Part Confidence Maps stage output?

A Raw input images

B Heatmaps showing where body parts are likely located

C Vector fields connecting body parts

D Final pose coordinates

Key Insight

OpenPose uses a two-step approach: first detecting body parts with confidence maps, then connecting them with affinity fields. This helps it accurately find human poses even in complex images.