Source-linked AI summary
Tiny SSD: A Tiny Single-shot Detection Deep Convolutional Neural Network for Real-time Embedded Object Detection
Alexander Wong, Mohammad Javad Shafiee, Francis Li, Brendan Chwyl
TL;DR
Embedded object detection requires accurate classification and localization without the high computational and memory demands of larger networks. Tiny SSD combines optimized Fire modules with SSD-based auxiliary layers and achieves a 2.3MB model, 571.09 million MAC operations, and 61.3% mAP on VOC 2007 while outperforming Tiny YOLO in the reported comparison.
Problem
Object-detection networks can be too computationally and memory intensive for embedded deployment, while efficient architectures must retain accuracy across varied object categories.
Method
Tiny SSD combines a non-uniform optimized Fire sub-network stack with optimized SSD-based auxiliary convolutional feature layers and predictors.
Results
2.3MB model size, 571.09 million MAC operations, and 61.3% mAP on VOC 2007; the model is ∼26X smaller and achieves ∼4.2% higher mAP than Tiny YOLO.
Takeaways & Limitations
The reported results support designing very small deep neural networks for real-time object detection in embedded scenarios.
Abstract
from arXiv · showhide
Object detection is a major challenge in computer vision, involving both object classification and object localization within a scene. While deep neural networks have been shown in recent years to yield very powerful techniques for tackling the challenge of object detection, one of the biggest challenges with enabling such object detection networks for widespread deployment on embedded devices is high computational and memory requirements. Recently, there has been an increasing focus in exploring small deep neural network architectures for object detection that are more suitable for embedded devices, such as Tiny YOLO and SqueezeDet. Inspired by the efficiency of the Fire microarchitecture introduced in SqueezeNet and the object detection performance of the single-shot detection macroarchitecture introduced in SSD, this paper introduces Tiny SSD, a single-shot detection deep convolutional neural network for real-time embedded object detection that is composed of a highly optimized, non-uniform Fire sub-network stack and a non-uniform sub-network stack of highly optimized SSD-based auxiliary convolutional feature layers designed specifically to minimize model size while maintaining object detection performance. The resulting Tiny SSD possess a model size of 2.3MB (~26X smaller than Tiny YOLO) while still achieving an mAP of 61.3% on VOC 2007 (~4.2% higher than Tiny YOLO). These experimental results show that very small deep neural network architectures can be designed for real-time object detection that are well-suited for embedded scenarios.
I. INTRODUCTION
Object detection networks can be accurate but too computationally and memory intensive for embedded deployment. Tiny SSD combines efficient Fire modules with SSD-based layers to target small models, real-time inference, and object-detection performance.
- Object detection requires both object classification and localization within a scene.
- Embedded deployment is constrained by high computational and memory requirements, with some networks achieving only single-digit frame rates on high-end GPUs.
- Existing compact detectors still present deployment or evaluation gaps, including large YOLO model sizes and limited demonstrated object categories for SqueezeDet.
- Tiny SSD combines a non-uniform optimized Fire sub-network stack with SSD-based auxiliary convolutional feature layers.
II. OPTIMIZED FIRE SUB-NETWORK STACK
Tiny SSD’s first sub-network stack uses a standard convolutional layer followed by optimized, non-uniform Fire modules. The design applies squeeze-and-expand operations while empirically selecting module count and configurations to balance size, speed, and detection performance.
- The first stack consists of a standard convolutional layer followed by highly optimized Fire modules.
- The Fire design reduces 3 × 3 filters and their input channels through a 1 × 1 squeeze layer followed by 1 × 1 and 3 × 3 expand filters.
- The optimized stack uses 10 Fire modules, selected empirically for strong object-detection performance.
- Each Fire module is configured to minimize parameters while maintaining overall detection accuracy, producing a non-uniform stack.
III. OPTIMIZED SUB-NETWORK STACK OF SSD-BASED CONVOLUTIONAL FEATURE LAYERS
Tiny SSD’s second stack adds optimized SSD-based auxiliary convolutional feature layers and predictors to the Fire stack. Their microarchitectures are configured to reduce parameters while preserving detection accuracy.
- The second stack uses auxiliary convolutional feature layers and convolutional predictors inspired by SSD’s multi-scale detection architecture.
- The auxiliary layers progressively decrease in size, providing flexibility for detecting objects at multiple scales.
- The architecture tables report the optimized Fire stack and auxiliary layers, including layer inputs, filters, kernel sizes, and parameters.
- The number of filters in each auxiliary feature layer is optimized to minimize parameters while preserving full-network detection accuracy.
IV. PARAMETER PRECISION OPTIMIZATION
Tiny SSD uses reduced parameter precision to further lower its computational and memory footprint for embedded scenarios.
- Tiny SSD represents parameters in half-precision floating-point to reduce model size for stricter embedded requirements.
- The precision reduction is reported to have a negligible effect on object-detection accuracy.
V. EXPERIMENTAL RESULTS AND DISCUSSION
The experiments evaluate Tiny SSD and Tiny YOLO on VOC2007/2012, using model size, detection accuracy, and computational operations as comparison criteria.
- Tiny YOLO serves as the baseline because it is a popular embedded detector with a 60.5MB model and 6.97 billion operations.
- VOC2007/2012 contain natural images annotated with 20 object categories for evaluating embedded object detection networks.
- Mean average precision is computed on the VOC2007 test dataset after training the networks on the VOC2007/2012 training datasets.
A. Training Setup
Tiny SSD was trained in Caffe for 220,000 iterations using batches of 24 and RMSProp optimization.
- 220,000 iterations were used to train Tiny SSD in the Caffe framework with a batch size of 24.
- RMSProp training used a base learning rate of 0.00001 and γ = 0.5.
B. Discussion
Tiny SSD substantially reduces model size while improving VOC2007 detection accuracy over Tiny YOLO, and its inference cost supports embedded deployment.
- 2.3MB is Tiny SSD’s model size, approximately 26X smaller than Tiny YOLO.
- 61.3% mAP is achieved by Tiny SSD on VOC2007, approximately 4.2% higher than Tiny YOLO.
- Figure 4 presents Pascal VOC examples with ground-truth bounding boxes and object categories.
- 571.09 million MAC operations are required for Tiny SSD inference, supporting its suitability for real-time embedded object detection.
VI. CONCLUSIONS
The paper introduces Tiny SSD as a compact single-shot detector that combines optimized Fire and SSD-based components for embedded object detection.
- Tiny SSD combines a non-uniform Fire sub-network stack with optimized SSD-based auxiliary convolutional feature layers.
- Tiny SSD is approximately 26X smaller than Tiny YOLO, requires 571.09 million MAC operations, and achieves approximately 4.2% higher VOC2007 mAP.
- The results support designing very small deep neural networks for real-time object detection in embedded scenarios.