SSD (Single Shot Detector)

In an AI glossary, SSD does not stand for a storage drive. It is the Single Shot MultiBox Detector: a neural network architecture for object detection introduced by Wei Liu and colleagues in the paper "SSD: Single Shot MultiBox Detector," published at ECCV 2016 and posted on arXiv in December 2015. Its contribution was to show that a single network, in a single pass over an image, could locate and classify objects about as accurately as the most precise systems of its day, at real-time speed.

Sharing an acronym with the Solid State Drive is pure coincidence: one is storage hardware, the other a milestone in computer vision.

How it works: one pass and default boxes

The dominant detectors of the time worked in two stages: one part of the network first proposed image regions likely to contain something, and another then classified each proposal. SSD drops that first stage. As the original paper by Liu and colleagues describes it, the network discretizes the space of possible bounding boxes into a set of default boxes with different aspect ratios and sizes, anchored to every position of several feature maps. For each box, the network simultaneously predicts a score for every category and a fine adjustment of its coordinates.

The second ingredient is multi-scale prediction: SSD combines outputs from feature maps at different resolutions, so earlier, higher-resolution layers capture small objects while deeper layers handle large ones. Everything happens in a single pass, hence "single shot."

Against Faster R-CNN and YOLO

In the paper's experiments, the SSD300 model reached 74.3% mAP (mean average precision) on the PASCAL VOC2007 benchmark while processing 59 images per second on a Titan X GPU. Faster R-CNN, the reference two-stage detector, delivered similar accuracy (around 73%) but only 7 images per second; the first version of YOLO, the other single-pass detector, was fast (45 images per second) but stopped at 63.4%. In 2016, SSD offered the best of both worlds, and its authors credited the edge over YOLO to default boxes at multiple scales.

Where SSD stands in 2026

At the research frontier, SSD has been surpassed. Today's detection leaderboards, such as those maintained by Roboflow and Ultralytics, are topped by transformer-based detectors descended from DETR — RT-DETR and RF-DETR, which do away with anchors and non-maximum suppression — and by recent members of the YOLO family, such as YOLO26, released by Ultralytics in January 2026.

Even so, SSD is not a museum piece. Paired with lightweight backbones such as MobileNet, it is still deployed on edge devices — Raspberry Pi, Jetson — through frameworks like TensorFlow Lite, and edge-computing benchmark studies keep including it among the models they evaluate. Its core idea, predicting over reference boxes at several scales in a single pass, shaped a good share of the detectors that came after it.