SSD

The paper makes object detection a single-shot prediction problem over default boxes placed on multiple feature maps.

Metadata

Field	Value
Paper	SSD: Single Shot MultiBox Detector
Authors	Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, Alexander C. Berg
Year	2016
Venue	ECCV 2016
arXiv	1512.02325
Status	seed note started

One-Line Takeaway

SSD removes proposal generation and predicts class scores plus box offsets for many default boxes across several feature-map resolutions.

Question

YOLO showed that object detection can be done in one forward pass, but early one-stage detectors struggled with localization and object scale.

SSD asks:

Can a single-shot detector improve scale and aspect-ratio coverage by predicting from multiple feature maps and multiple default boxes?

The interface is:

$$I\to \{F_l{\}}_{l=1}^L\to \{({\hat{p}}_{i,a},{\hat{t}}_{i,a})\}.$$

where:

• $F_l$ is a feature map at level $l$;
• $i$ is a spatial location;
• $a$ is a default box index;
• $\hat{p}$ is a class distribution;
• $\hat{t}$ is a box offset.

Architecture Contract

Component	Input	Output	Role
base CNN	image	feature hierarchy	visual representation
extra feature layers	deep features	lower-resolution maps	larger receptive fields
default boxes	feature locations	anchor-like priors	cover scales/aspect ratios
prediction heads	each feature map	class logits and box offsets	dense one-stage detection
NMS	candidate detections	final detections	remove duplicate boxes

The core architecture:

$$\text{image}\to \text{multi-scale feature maps}\to \text{default-box predictions}.$$

Default Boxes

At each feature-map cell, SSD defines several default boxes:

$$D_{l,i}=\{d_{l,i,1},\dots ,d_{l,i,A_l}\}.$$

Each default box has a scale and aspect ratio. The model predicts offsets:

$${\hat{t}}_{l,i,a}=(\Delta x,\Delta y,\Delta w,\Delta h),$$

and class logits:

$${\hat{p}}_{l,i,a}\in {\mathbb{R}}^{C+1}.$$

The final box is:

$${\hat{b}}_{l,i,a}=\mathrm{decode}(d_{l,i,a},{\hat{t}}_{l,i,a}).$$

The background class accounts for default boxes that do not match objects.

Multi-Scale Prediction

SSD predicts from several feature maps:

$$\{F_1,F_2,\dots ,F_L\}.$$

Higher-resolution maps help smaller objects:

$$F_{\text{early}}\to \text{small-object candidates}.$$

Lower-resolution maps help larger objects:

$$F_{\text{late}}\to \text{large-object candidates}.$$

This is a key difference from a detector that predicts only from one final feature grid.

Training Objective

SSD matches ground-truth boxes to default boxes and trains classification plus localization:

$$\mathcal{L}=\frac{1}{N}\left({\mathcal{L}}_{\mathrm{conf}}+\alpha {\mathcal{L}}_{\mathrm{loc}}\right).$$

The localization term is usually Smooth L1 over matched boxes:

$${\mathcal{L}}_{\mathrm{loc}}=\sum _{(i,j)\in \mathcal{M}}SmoothL1\left(t_i-t_j^{\text{*}}\right),$$

where $\mathcal{M}$ is the matching between default boxes and ground-truth boxes.

Why It Matters

SSD is important because it makes the one-stage detector design more general than the original YOLO grid:

$$\text{single-shot}+\text{default boxes}+\text{multi-feature-map prediction}.$$

This pattern directly anticipates later dense detectors and mobile detection heads.

Comparison

Axis	YOLO	SSD
Prediction locations	grid cells	multiple feature maps
Box priors	grid responsibility	default boxes per location
Scale handling	limited in original YOLO	explicit multi-level prediction
Detector type	one-stage	one-stage
Main tradeoff	speed and simplicity	more anchors/default boxes to tune

What To Watch

• Default-box scale/aspect design strongly affects recall.
• Hard negative mining is part of the training recipe, not only the architecture.
• Multi-scale feature prediction is not the same as the later FPN top-down fusion.
• NMS remains part of inference.
• Later SSD variants differ through backbones, feature pyramids, and mobile heads.

Related

• Object detection
• CNN
• YOLO
• Feature Pyramid Networks
• RetinaNet
• Faster R-CNN