1. Efficient in Object Detection

1-1. Model Scaling

• Width scaling
• Depth scaling
• Resolution Scaling

📌Can we scale up a model to gain better accuracy and efficiency?

→ There is an optimal way to scale up width, depth and resolution.

2. EfficientNet

• The need for fast and small models $\uparrow$
• Most models trade-off efficiency and accuracy (MobileNets, SqueezeNets)
• EfficientNet gains efficiency in big SOTA models by optimizing scaling up

2-1. Scale up

• Width scaling
  • Wide networks catch fine features and are easier to train
  • But are week at catching high level features

• Depth scaling
  • Deep networks catch high-level features well and are good at generalization
  • But gradient vanishing problem makes them harder to train

• Resolution scaling
  • Good at catching fine features

2-2. Accuracy & Efficiency

• The growing rate of accuracy decreases if we keep scaling up by one factor