A Comparative Study of Real-Time Object Detection Using YOLOv8 and SSD-MobileNet
Ali Hamza (Self-directed mini-research)
INTRODUCTION
Real-time object detection is essential for surveillance, traffic monitoring, and smart automation systems. This study compares YOLOv8 and SSD-MobileNet under CPU-only constraints, where deployment cost and hardware limits are major factors.
RESEARCH STATEMENT
Selecting a detector for limited hardware is challenging because developers must balance three factors: inference speed, localization quality, and ease of deployment.
OBJECTIVE
The objective is to evaluate both models on the same test conditions and determine which model is better for (1) accuracy-focused projects and (2) rapid deployment projects.
METHODOLOGY
- We used a Pascal VOC subset with three classes: Person, Car, and Bicycle.
- YOLOv8 was fine-tuned using transfer learning.
- SSD-MobileNet was used with pretrained weights (no additional training).
- Both models were evaluated on the same validation images.
- Inference latency and qualitative bounding-box performance were compared.
EXPERIMENTAL SETUP
Evaluation was conducted on an Intel Core i5 machine using Python 3.11 with CPU-only builds of PyTorch and TensorFlow. Image size was standardized, and each sample was run multiple times to compute stable average latency.
RESULTS AND ANALYSIS
Average inference time was nearly identical: YOLOv8 ~70 ms/image and SSD-MobileNet ~69.37 ms/image (around 14 FPS on CPU). YOLOv8 delivered tighter and more consistent bounding boxes, especially in overlapping or small-object cases. SSD-MobileNet showed slightly looser localization but required minimal setup and no training pipeline.
CONCLUSION
Both models are suitable for CPU deployment. YOLOv8 is preferable when detection precision is critical and training effort is acceptable, while SSD-MobileNet is better when immediate deployment and low implementation overhead are the priority.