Course contentsShow
← Machine Learning and Deep Learning
Lesson 968 of 3,53822. Computer Vision TasksFree lesson

SSD: Multi-Scale Feature Maps for Detection

How Single Shot Detector uses multiple convolutional layers to detect objects at different scales.

SSD: Multi-Scale Feature Maps for Detection

What You'll Learn

You'll discover how Single Shot Detector (SSD) detects objects of different sizes by extracting features from multiple layers of a CNN simultaneously.

The Core Problem

Imagine trying to spot both a distant airplane and a nearby car in the same photo. Your eyes naturally focus differently for each—you need wide context for small objects and fine detail for large ones. Object detectors face the same challenge: how to find tiny and huge objects in one pass?

How SSD Solves It

SSD uses multiple convolutional layers at different depths as detection layers. Unlike YOLO which predicts from a single feature map, SSD makes predictions from several layers:

  • Early layers (higher resolution, smaller receptive fields) → detect small objects
  • Deep layers (lower resolution, larger receptive fields) → detect large objects

Think of it like having multiple magnifying glasses with different zoom levels, all working simultaneously.

The Multi-Scale Architecture

SSD starts with a base CNN (like VGG) and adds extra convolutional layers that progressively decrease in spatial size. At each chosen layer, it:

  1. Applies small 3×3 convolutions to predict bounding boxes and class scores
  2. Uses default boxes (anchor boxes) of various aspect ratios at each scale
  3. Combines predictions from all layers during training and inference

For example, a 38×38 feature map might detect small faces, while an 8×8 feature map detects large vehicles—all in one forward pass.

Why It Matters

This multi-scale approach gives SSD a sweet spot: faster than two-stage detectors like R-CNN (since it's one-stage) while more accurate than early YOLO (thanks to multi-scale features). It naturally handles objects at vastly different sizes without needing separate passes.

Key Takeaway: SSD achieves efficient multi-scale object detection by making predictions from multiple convolutional feature maps at different resolutions, allowing it to detect both small and large objects in a single forward pass.