YOLO

The paper frames object detection as a single regression problem from a full image to boxes and class probabilities.

Metadata

Field	Value
Paper	You Only Look Once: Unified, Real-Time Object Detection
Authors	Joseph Redmon, Santosh Divvala, Ross Girshick, Ali Farhadi
Year	2016
Venue	CVPR 2016
arXiv	1506.02640
Paper	CVF Open Access PDF
Status	seed note started

One-Line Takeaway

YOLO replaces proposal-based detection with a single CNN that predicts bounding boxes and class probabilities directly from the full image in one forward pass.

Question

Earlier detection pipelines often looked like:

$$\text{region proposals}\to \text{classification}\to \text{box refinement}\to \text{post-processing}.$$

YOLO asks:

Can object detection be trained as one direct prediction problem?

The proposed interface:

$$I\to \hat{Y},$$

where:

$$\hat{Y}\in {\mathbb{R}}^{S\times S\times (B\cdot 5+C)}.$$

Here:

• $S\times S$ is the image grid;
• $B$ is the number of boxes per grid cell;
• $C$ is the number of classes;
• each box predicts $(x,y,w,h,\text{confidence})$.

Architecture Contract

Component	Input	Output	Role
CNN backbone	image	spatial feature grid	full-image visual representation
grid predictor	feature grid	box and class predictions per cell	single-stage dense prediction
detection loss	predicted grid and targets	scalar objective	train boxes, objectness, and classes together
NMS/post-process	candidate boxes	final detections	remove duplicate boxes

The main architectural shift:

$$\text{proposal-classification pipeline}\to \text{single full-image predictor}.$$

Grid Prediction

YOLO divides the image into an $S\times S$ grid. Each grid cell is responsible for objects whose center falls inside that cell.

For cell $(i,j)$ and box $b$, the model predicts:

$${\hat{y}}_{i,j,b}=(\hat{x},\hat{y},\hat{w},\hat{h},\hat{c}).$$

The confidence combines objectness and localization quality:

$$\hat{c}\approx P(\text{object})\cdot \mathrm{IoU}(\hat{b},b^{\text{*}}).$$

Class probabilities are predicted per cell:

$${\hat{p}}_{i,j,k}=P(c=k\mid \text{object in cell }i,j).$$

The final class-specific score is:

$$s_{i,j,b,k}={\hat{c}}_{i,j,b}\cdot {\hat{p}}_{i,j,k}.$$

Unified Detection Loss

A simplified YOLO-style objective combines localization, confidence, and classification:

$$\mathcal{L}={\lambda }_{\mathrm{coord}}{\mathcal{L}}_{\mathrm{box}}+{\mathcal{L}}_{\mathrm{conf}}+{\mathcal{L}}_{\mathrm{cls}}.$$

The original formulation weights coordinate error and no-object confidence differently:

$${\mathcal{L}}_{\mathrm{box}}=\sum _{i,j,b}{1}_{i,j,b}^{\mathrm{obj}}\left[(x-\hat{x}{)}^2+(y-\hat{y}{)}^2+(\sqrt{w}-\sqrt{\hat{w}}{)}^2+(\sqrt{h}-\sqrt{\hat{h}}{)}^2\right].$$

The square-root width/height terms reduce the penalty imbalance between small and large boxes.

Why It Matters

YOLO is the canonical one-stage real-time detector starting point:

$$\text{one image}\to \text{one network evaluation}\to \text{detections}.$$

It is architecturally important because it makes speed and simplicity part of the model contract. The design sacrifices some localization precision compared with two-stage detectors, but it makes detection usable in real-time systems.

Comparison

Axis	Faster R-CNN	YOLO
Detection style	two-stage	one-stage
Candidate generation	RPN proposals	grid predictions
Speed target	accurate proposal detector	real-time detection
Context	region-centric	full-image
Main weakness	heavier pipeline	localization errors, small objects

Reading Boundary

YOLO has many later versions. This note is about the original YOLOv1 architecture paper. Later YOLO variants add anchors, stronger backbones, feature pyramids, better assignment rules, and many training changes. Those should not be retroactively attributed to the original paper.

What To Watch

• Grid-cell responsibility can struggle when multiple small objects fall in the same cell.
• Direct regression is fast but can trade away localization precision.
• The final detector still uses post-processing such as non-maximum suppression.
• Later YOLO versions differ substantially from YOLOv1.
• Benchmark comparisons should separate speed, input resolution, backbone, and dataset.

Related

• Object detection
• CNN
• SSD
• Feature Pyramid Networks
• Faster R-CNN
• DETR