EMOv2: Pushing 5M Vision Model Frontier.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Lightweight models are crucial for efficient deployment on mobile devices with limited resources.
• Existing lightweight Convolutional Neural Network (CNN) architectures like Inverted Residual Blocks (IRBs) lack attention-based counterparts.
• Transformers offer attention mechanisms but can be computationally intensive for lightweight applications.

Purpose of the Study:

• To develop parameter-efficient and lightweight models for dense prediction tasks.
• To establish a new performance frontier for models in the 5 million parameter magnitude range.
• To unify the design principles of efficient IRBs and Transformer components.

Main Methods:

• Reimagined lightweight infrastructure by unifying efficient IRBs and Transformer components.
• Abstracted a one-residual Meta Mobile Block (MMBlock) for lightweight model design.
• Developed an Improved Inverted Residual Mobile Block (i222222222222222222222222rmb) and a hierarchical Efficient MOdel version 2 (EMOv2).

Main Results:

• EMOv2 models (1M, 2M, 5M parameters) achieved 72.3, 75.8, and 79.4 Top-1 accuracy, significantly outperforming comparable CNN and attention-based models.
• EMOv2-5M with RetinaNet achieved 41.5 mAP for object detection, a +2.6 improvement over previous models.
• With an enhanced training recipe, EMOv2-5M reached 82.9 Top-1 accuracy, setting a new benchmark for 5M magnitude models.

Conclusions:

• EMOv2 demonstrates superior performance and efficiency for lightweight models across various vision tasks.
• The unified design approach effectively bridges CNN and attention-based architectures.
• EMOv2 pushes the performance limits for models in the 5 million parameter range, enabling advanced capabilities on mobile devices.