Object Detection
(2D)

Object detection is the task of finding objects inside an image, identifying what they are, and locating exactly where they appear. It is one of the foundations of modern computer vision.

What Is Object Detection?

Classification tells us what an image contains. Detection goes further: it tells us what objects exist, where they are, and how confident the model is about each prediction.

For example, a single image may contain three people, two cars, and a bicycle. A detector must find every object separately and assign each one a class.

Bounding Boxes: The Language of Detection

The most common output of 2D object detection is a bounding box. A bounding box is a rectangle that surrounds a detected object.

What Does a Detection Contain?

Detection:
{
    class: "car",
    confidence: 0.94,
    box: [120, 80, 520, 410]
}

Intersection over Union (IoU)

To determine whether a predicted box is correct, we compare it with the ground-truth box using IoU.

                    Area of Overlap
IoU = ─────────────────────────────────────
       Area of Prediction ∪ Ground Truth

IoU = 1.0  → Perfect overlap
IoU = 0.0  → No overlap

A higher IoU means that the predicted box is closer to the real location of the object.

CNN-Based Object Detection

Before transformers became dominant, CNNs were the main engine behind modern object detection.

A Convolutional Neural Network (CNN) learns visual features by applying convolutional filters across the image. Early layers learn simple patterns, while deeper layers learn more meaningful visual structures.

How CNN Features Build Up

From CNN Features to Detection

A detector uses the feature maps produced by a CNN backbone and predicts where objects are located and what classes they belong to.

Input Image
     ↓
CNN Backbone
     ↓
Feature Maps
     ↓
Detection Head
     ↓
Bounding Boxes + Classes + Confidence

This CNN-based design is the foundation of many highly successful detectors, including the YOLO family.

Transformer-Based Object Detection

Transformers introduced a different way of understanding images: instead of processing only local visual patterns, they can model relationships between different regions of the image.

Vision Transformers divide an image into patches and use self-attention to determine which parts of the image are important to each other.

Why Attention Matters

Imagine an image containing a person holding a tennis racket. Understanding the relationship between the person, the hand, and the racket can be useful for detection and visual reasoning.

Image Regions:

[ Person ] ←──── Attention ────→ [ Racket ]
     ↑                                  ↑
     └──────── Global Context ─────────┘

Transformer-based detectors can use this global context to reason about the image as a whole rather than treating every region independently.

YOLO: Fast Detection with CNNs and Modern Backbones

YOLO — You Only Look Once is one of the most recognizable names in object detection. The goal is simple: process an image efficiently and predict detections in a single model pipeline.

The Main Parts of a YOLO Detector

YOLO models are widely used because they provide a practical balance between accuracy, speed, and ease of deployment. This makes them useful for real-time applications such as cameras, robotics, industrial inspection, and edge AI.

RF-DETR: Transformer-Based Detection

RF-DETR represents a modern DETR-style approach to object detection. Instead of following the traditional YOLO-style detection design, it uses transformer-based visual reasoning.

Object Queries

A key idea in DETR-style models is to predict a fixed set of possible objects. Each prediction can represent one object in the image.

Image
  ↓
Transformer Encoder
  ↓
Object Queries

Query 1 → Person
Query 2 → Car
Query 3 → Bicycle
Query 4 → No Object

This provides a fundamentally different perspective on detection compared with traditional pipelines that generate many candidate boxes and then filter duplicates.

CNN Detectors vs Transformer Detectors

Object Detection in Vyrion Vision

Use supported object detection models such as YOLO and RF-DETR to perform inference on images and other supported inputs.

Upload a trained model and test it on new images. Inspect bounding boxes, classes, confidence scores, and failure cases.

Run different models on the same input and compare the results. For example, you can compare a YOLO detector with an RF-DETR detector on the same image or dataset.

Same Image
     │
     ├──→ YOLO
     │       └──→ Boxes + Classes + Confidence
     │
     └──→ RF-DETR
             └──→ Boxes + Classes + Confidence

Compare:
• Which objects were detected?
• Which boxes are more accurate?
• Which model misses objects?
• Which model produces false detections?
• Which model is faster?

That is the difference between simply downloading a model and actually understanding how models perform on your data.

How Detection Models Are Evaluated

A good object detector must locate objects accurately and identify them correctly.

Real-World Model Evaluation:

Accuracy
    +
Speed
    +
Memory
    +
Robustness
    +
Deployment Requirements

Key Takeaways