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

What problem it addressed

Object detection asks a computer not only what is in an image but where each object is. A common strategy in prior work was to take classifiers trained to recognize objects and repurpose them for detection. The YOLO paper (You Only Look Once: Unified, Real-Time Object Detection) was submitted to arXiv on June 8, 2015 by Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. It proposed a different starting point: treat detection itself as one prediction problem rather than a pipeline organized around classifiers.^[1]

How it works

Rather than repurposing classifiers, the paper frames object detection as a regression problem over spatially separated bounding boxes and their associated class probabilities. A single neural network takes a full image and predicts those outputs directly in one evaluation. The model therefore learns to produce locations and class labels together, instead of using a classifier's outputs as the detection mechanism. Because the whole detection pipeline is a single network, it can be optimized end to end directly on detection performance.^[1]

Results

The base YOLO model processes images in real time at 45 frames per second. A smaller version, Fast YOLO, processes 155 frames per second while still achieving double the mean average precision (mAP) of other real-time detectors. These figures show a single-network detector operating at high speed without giving up accuracy relative to other real-time systems.^[1]

The paper also reports a trade-off. Compared with state-of-the-art detection systems, YOLO makes more localization errors: its predicted boxes are less precise about an object's exact extent. But it is far less likely to predict false detections where nothing exists. This matters because false detections on empty background can be costly in practical use. The authors further test generalization from natural images to artwork. On the Picasso and People-Art datasets, YOLO outperforms DPM and R-CNN by a wide margin.^[1]

Why it matters

The paper's importance lies in its framing. Detection becomes a single regression task over boxes and class probabilities, and the entire pipeline becomes one network that can be trained end to end on the detection objective.^[1] That unification is tied to speed: 45 frames per second for the base model and 155 for Fast YOLO.^[1] It also changes the error profile. YOLO is less accurate at localization than state-of-the-art systems, but much less prone to false detections on empty background.^[1] The artwork results suggest that the network learns object representations general enough to survive a shift in visual domain.^[1] Taken together, the results present real-time object detection as a problem of direct prediction rather than a chain of separately tuned classifiers.^[1]