Learn→Computer Vision Systems→Edge Deployment (MobileNet, EfficientNet-Lite)→6 of 6

Computer Vision Systems • Edge Deployment (MobileNet, EfficientNet-Lite)Hard⏱️ ~3 min

Accuracy vs Latency Trade-offs: Choosing Between SSD MobileNet and EfficientDet Lite

THE PARETO FRONTIER
Every edge model sits on an accuracy-latency trade-off curve. MobileNetV2 at 5ms achieves 72% ImageNet accuracy. EfficientNet-Lite4 at 30ms achieves 80%. There is no model that is simultaneously fastest and most accurate. Your job is to find the point on this curve that meets your application requirements.
DETECTION MODEL COMPARISON
SSD MobileNetV2: 15-25ms inference, 22% COCO mAP. Good for real-time when accuracy is less critical (presence detection, counting).
EfficientDet-Lite0: 30-40ms, 26% mAP. Better accuracy, still real-time capable.
EfficientDet-Lite2: 60-80ms, 32% mAP. Near-cloud accuracy but sacrifices real-time.
YOLOv5n: 20-30ms, 28% mAP. Good balance for many applications.
DECISION FRAMEWORK
Start with your latency budget, then find the best accuracy within it. For 30fps video, budget 33ms. Subtract preprocessing (3ms) and postprocessing (5ms). You have 25ms for inference. EfficientDet-Lite0 or SSD-MobileNet fits; EfficientDet-Lite2 does not.
💡 Key Insight: Do not optimize for benchmark accuracy. Optimize for task accuracy within your latency budget. A model that is 5% more accurate but misses frames is worse than a faster model that processes every frame.
WHEN TO SACRIFICE ACCURACY
Prefer lower accuracy when: (1) task tolerates errors (suggestions vs safety-critical). (2) downstream processing can correct mistakes. (3) real-time response is essential. Prefer higher accuracy when: (1) errors are costly. (2) batch processing is acceptable. (3) thermal headroom exists.

💡 Key Takeaways

✓Pareto frontier: no model is fastest AND most accurate; find the point meeting your requirements

✓Detection comparison: SSD-MobileNet (25ms, 22% mAP), EfficientDet-Lite0 (35ms, 26%), YOLO5n (25ms, 28%)

✓Decision framework: start with latency budget, subtract pre/post processing, find best accuracy in remaining time

✓Task accuracy matters: 5% more accurate but missing frames is worse than faster with every frame

📌 Interview Tips

1Walk through the decision framework: 33ms budget - 3ms preprocess - 5ms postprocess = 25ms for inference

2Compare detection models with latency and mAP numbers for different use cases

3Emphasize task accuracy vs benchmark accuracy: processing every frame may beat higher accuracy with drops

← Back to Edge Deployment (MobileNet, EfficientNet-Lite) Overview