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

A detection problem framed as regression

On 8 June 2015, Joseph Redmon, Santosh Divvala, Ross Girshick and Ali Farhadi posted "You Only Look Once: Unified, Real-Time Object Detection" to arXiv ^[1]. The paper addresses object detection: given an image, a system must say where objects are and which class each object belongs to. In the paper's formulation, the answer is a set of bounding boxes—rectangles around objects—together with class probabilities. The authors framed this task as a single regression problem, straight from image pixels to bounding box coordinates and class probabilities ^[1]. A regression problem, in this context, means the network learns to output numbers—coordinates and probabilities—directly. The key move is to treat the whole detection output as something a network can predict directly.

One network, one evaluation

The paper described a single neural network that predicts bounding boxes and class probabilities directly from full images in one evaluation ^[1]. The phrase "one evaluation" is central: the image goes through the network once, and the network produces its detection outputs. The architecture is unified, and the authors said it could be trained end-to-end directly on detection performance ^[1]. End-to-end training means the network's parameters are adjusted according to the detection objective itself. The paper therefore presents detection as a single learned mapping from a full image to boxes and class probabilities.

Speed and error profile

The reported speeds made the approach notable. The base YOLO model processed images in real time at 45 frames per second ^[1]. A smaller version, Fast YOLO, processed 155 frames per second ^[1]. Frames per second is a measure of throughput: how many images the system can process in one second. These numbers matter because object detection is often discussed in terms of accuracy alone, but many uses also depend on whether a system can keep up with a stream of images. The paper's results gave a clear picture of that speed.

The paper also reported a trade-off in errors. YOLO made more than twice as many localization errors as the best detection methods of the time, but it was less likely to predict false detections in the background ^[1]. A localization error means the predicted box is in the wrong place or has the wrong extent; a false detection in the background means the system reports an object where there is none. The reported pattern is therefore not simply "better" or "worse": the method was weaker at precise localization and stronger at avoiding background false positives, according to the paper. This is a useful way to describe a detector's behavior, because different applications care about different kinds of mistakes.

What the paper established

The paper's contribution is the combination of formulation and engineering. It framed object detection as one regression problem, implemented it with a single neural network, and trained the unified architecture end-to-end on detection performance ^[1]. It reported real-time speed for the base model and an even faster smaller model ^[1]. It also reported an error analysis: more localization errors than the best methods of the time, but less likely to produce background false detections ^[1]. The title, "You Only Look Once," names the central idea—one look, one evaluation—and the paper supplies the concrete version: full images in, bounding boxes and class probabilities out. For readers learning detection, the paper is a clear example of how a task can be re-described so that one network can learn it directly.