- Published on
Single Shot Detection
- Authors
- Name
- Vincent Hu
Introduction
The project implements the Single Shot Detection (SSD) algorithm, originally introduced in the paper "SSD: Single Shot MultiBox Detector". SSD is a real-time object detection model that balances speed and accuracy, making it effective for applications like autonomous driving and real-time surveillance.
Overview of SSD
Unlike two-stage detectors like Faster R-CNN, SSD is a single-stage detector, meaning it directly predicts object classes and bounding boxes in one forward pass of the network. This efficiency comes from its multi-scale feature maps and convolutional predictors, allowing it to detect objects at different scales with minimal computational overhead.
SSD consists of a base network (e.g., VGG16) for feature extraction and additional convolutional layers for multi-scale detection. The key aspects of SSD include:
- Multi-scale feature maps: Predictions are made at different layers, detecting both small and large objects.
- Default boxes (anchor boxes): Predefined boxes at various aspect ratios help detect objects of different shapes efficiently.
- Hard negative mining: Balances positive and negative samples to stabilize training.
- Smooth L1 and cross-entropy loss: Optimizes bounding box regression and classification simultaneously.
Implementation Details
The implementation follows the standard SSD pipeline:
- Backbone Network: A modified VGG16 extracts feature maps.
- Multi-scale Detection: Additional convolutional layers process these maps.
- Anchor Matching: Ground truth boxes are assigned to predefined anchors.
- Loss Computation: Classification (cross-entropy) and localization (Smooth L1) losses are computed.
- Training with Augmentations: Data augmentation (random cropping, flipping, color distortions) improves robustness.
Conclusion
SSD offers a fast and efficient solution for real-time object detection. While it may not match the accuracy of transformer-based models like DETR, it remains a solid choice for resource-constrained environments. This implementation reproduces SSD with modifications for improved performance and flexibility.