What is the Mixture of Gaussians(MoG)

 

🎯 Context First: Why Learn Part Positions?

When you’re trying to detect parts of an object (like a bird’s head or tail), it helps to know where those parts typically appear relative to the object.

• For example: The head is usually above the body.

• But because of pose variations, one single "mean" location isn’t enough.

That’s why the model uses a Mixture of Gaussians — to capture multiple common configurations of part locations.

---

🧠 What is a Mixture of Gaussians (MoG)?

A Mixture of Gaussians is a probabilistic model that represents a complex distribution as a weighted sum of multiple Gaussian (bell curve) distributions.

Mathematically:

$$P(x) = \sum_{k=1}^{K} \pi_k \cdot \mathcal{N}(x | \mu_k, \Sigma_k)$$

Where:

• $\pi_k$ = weight (mixing coefficient) of the $k$-th Gaussian (sums to 1)

• $\mu_k$ = mean (center) of the $k$-th Gaussian

• $\Sigma_k$ = covariance (spread/shape) of the $k$-th Gaussian

• $K$ = number of mixture components

---

🧭 How is it used to model part positions?

In Part-based R-CNN, here's how it works:

---

1️⃣ Collect Part Coordinates from Training Data

For each part type (e.g., head), gather the coordinates from annotated training images, relative to the whole object box.

Example:
If the bird’s body is at (x, y, w, h), and the head is at (xₕ, yₕ), then part coordinates are normalized as offsets.

---

2️⃣ Fit a Mixture of Gaussians to These Coordinates

Use the training data to fit a K-component MoG to these relative positions.

Each Gaussian captures a common configuration:

• One Gaussian might represent the bird facing left.

• Another could capture the bird flying, where the head is farther from the body.

📌 You can fit this using Expectation-Maximization (EM) algorithm.

---

3️⃣ Use the MoG as a Geometric Prior

At test time:

• You detect the object (e.g., bird body).

• Then, for each possible part location (e.g., a proposed head), you compute the likelihood of that position under the MoG model.

This becomes the δMG(x) scoring function in the paper — higher for "likely" locations.

So, in the paper’s scoring formula:

$$\Delta_{\text{geometric}}(X) = \Delta_{\text{box}}(X) \cdot \left( \prod_{i=1}^{n} \delta_i(x_i) \right)^\alpha$$

• $\delta_i(x_i)$ is the MoG likelihood of part $i$’s location.

• $\alpha$ is a weighting factor.

---

🔍 Example: Bird Head Detection

Let’s say in training data:

• Sometimes the head is 20 pixels above the center of the bird (Gaussian 1),

• Sometimes it’s 15 pixels to the left and 5 above (Gaussian 2),

• Sometimes it’s farther away (flying posture — Gaussian 3).

These patterns are captured as 3 Gaussians in the MoG.

At test time, a proposed head location gets a high score only if it aligns with one of these common patterns.

---

✅ Benefits of Using MoG

Advantage	Description
Multi-pose modeling	Handles pose variability with multiple Gaussians
Probabilistic & smooth	Assigns soft likelihoods instead of hard cutoffs
Lightweight computation	Gaussians are fast to evaluate
Easy to fit	Uses standard EM algorithm