Object Detection

Detecting Objects in Images

Object Detection is a fundamental task in Computer Vision that goes beyond simple image classification. While classification merely states “there is a dog in the image”, detection states “there is a dog here and a cat there”, drawing bounding boxes.

⚙️ How it works: Practical Context

The task involves two simultaneous predictions:

Regression: Predicting the coordinates of the bounding box (center x, center y, width, height).
Classification: Predicting the class (label) of the object inside that box.

There are two major families:

Two-Stage Detectors: (e.g., Faster R-CNN) Propose regions (RPN) and then classify/refine them using roi-pooling. Slower but highly accurate.
One-Stage Detectors: (e.g., YOLO, SSD) Treat detection as a single regression problem on a grid. Extremely fast, but with slight difficulty regarding small and clustered objects.

🧮 Mathematical Foundation (Multi-task Loss Function)

Because the model needs to optimize two distinct things at the same time, the total loss function ( $L_{total}$ ) is a linear combination of the classification loss ( $L_{cls}$ ) and the bounding box regression loss ( $L_{loc}$ ):

L_{total}(p, u, t, v) = L_{cls}(p, u) + \lambda \cdot [u \ge 1] L_{loc}(t, v)

Eq. 1: Multi-task Loss function.

Where:

$p$ is the predicted probability distribution (calculated via Softmax).
$u$ is the true object class ( $u \ge 1$ means it’s an object, $u=0$ indicates background).
$t$ is the predicted coordinate vector $(t_x, t_y, t_w, t_h)$ .
$v$ is the true ground truth bounding box vector.
$\lambda$ is the balancing hyperparameter.

1. Classification Loss (Cross-Entropy)

Usually computed with standard logarithmic cross-entropy over $K+1$ classes (where $+1$ is the background):

L_{cls}(p, u) = -\log p_u

Eq. 2: Cross-entropy loss for classification.

Where:

$L_{cls}(p, u)$ : classification loss for the object.
$p_u$ : probability predicted by the model for the correct class $u$ .

2. Regression Loss (Smooth L1 / Huber Loss)

For localization, using Mean Squared Error ( $L_2$ ) can cause gradients to explode on early outliers. Thus, we use the Smooth $L_1$ Loss, which is quadratic for small errors (smoothing convergence) and linear for large errors (outlier resistance):

L_{loc}(t, v) = \sum_{i \in \{x,y,w,h\}} \text{smooth}_{L_1}(t_i - v_i)

Eq. 3: Smooth L1 regression loss for bounding box.

Where:

\text{smooth}_{L_1}(x) = \begin{cases} 0.5 x^2 & \text{if } |x| < 1 \\ |x| - 0.5 & \text{otherwise} \end{cases}

🎯 When to use

Surveillance and tracking systems.
Autonomous vehicles (detecting pedestrians, signs).
Medical image analysis (identifying tumors in scans).