← AI Terminology
SSD - Single Shot Detector
SSD (Single Shot MultiBox Detector) is a one-stage object detection architecture that predicts bounding boxes and class labels directly from feature maps at multiple scales in a single forward pass — achieving real-time detection speed without a separate region proposal stage.
Introduced by Liu et al. (2016); one of the first practical real-time detectors.
Introduced by Liu et al. (2016); one of the first practical real-time detectors.
Why It Matters in AI
Two-stage detectors (Faster R-CNN) are accurate but slow — the region proposal network adds latency. SSD demonstrated that high-accuracy detection was possible with a single forward pass: predict bounding boxes from multiple feature map scales simultaneously, without region proposals. At 74% mAP on VOC2007 and 59 FPS on GPU, SSD set the template for one-stage detection. YOLO, RetinaNet, and modern one-stage detectors all build on this "predict from multi-scale features in one pass" principle.
Key Points
| Aspect | Description |
|---|---|
| vs YOLO | SSD explicitly uses multi-scale features; YOLO v1 used single scale — SSD better at small objects |
| Multi-scale | Extract feature maps at 6 different resolutions — detect small objects at high-res, large at low-res |
| Single pass | No region proposals — classification and box regression happen directly on feature maps |
| Anchor boxes | Predefined boxes at each location and scale — predict offsets from these anchors |
| Base network | VGG-16 backbone (original); updated versions use ResNet, MobileNet — trade accuracy for speed |
| Modern successors | YOLOv5/8, RetinaNet, EfficientDet — all build on SSD's one-stage multi-scale principle |
Simple Analogy
Looking at a map with multiple zoom levels simultaneously: a city map shows large regions; a street map shows small details. SSD examines the image at multiple resolutions at once — big objects found at coarse scale, small objects at fine scale — all in one look rather than first flagging regions of interest, then examining them.
Common Usage Examples
from torchvision.models.detection import ssd300_vgg16; model = ssd300_vgg16(pretrained=True)- MobileNet-SSD: lightweight variant for edge inference —
ssdlite320_mobilenet_v3_large - TensorFlow Object Detection API:
ssd_mobilenet_v2_320x320_coco17— Google's SSD implementation - OpenCV DNN:
net = cv2.dnn.readNetFromCaffe("deploy.prototxt", "res10.caffemodel")— SSD face detector - PASCAL VOC SSD300: 74.3% mAP @ 59 FPS on Titan X — original paper benchmark
Summary
In short: SSD detects objects by predicting bounding boxes at multiple feature map scales in a single forward pass — the foundational one-stage detector architecture that proved real-time high-accuracy detection is possible without region proposal networks.