You Only Look Once: Unified, Real-Time Object Detection

Inside the paper

The problem

Object detection pipelines often examined many candidate regions and then refined their predictions in separate stages. That made accurate detection costly when an application needed to process a stream of images.

The idea

YOLO treats detection as a direct prediction from the whole image to object locations and classes. One network produces the candidate detections together, allowing it to use image-wide context.

How it works

• The image is divided into a grid, and the cell containing an object’s center is responsible for predicting that object.
• Each cell predicts bounding boxes and confidence values, alongside probabilities for the object classes represented in that cell.
• Training combines errors in box coordinates, object confidence, and class predictions in a single loss, with different weights for different terms.
• At inference, class scores and box confidence are combined, and overlapping detections can be suppressed to remove duplicates.

What it showed

The paper reports real-time detection with its base model and analyzes the tradeoff against more accurate detectors. YOLO makes fewer background errors than Fast R-CNN in its analysis, but has more difficulty locating objects precisely.

Read it for

Figure 2 and Section 2 explain what the network predicts; Section 2.4 and Section 4.2 make the limitations concrete. You can skim the network training details in Section 2.2 before returning to the comparisons.

Written in our own words from the paper itself. A draft the editors have not yet reviewed.