Glance and Focus Networks for Dynamic Visual Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Visual recognition tasks often contain spatial redundancy, where only a subset of pixels holds discriminative information.
• Static models process all pixels equally, leading to inefficient time and space consumption.

Purpose of the Study:

• To develop an efficient visual recognition model that mimics the human visual system's coarse-to-fine processing.
• To reduce computational redundancy in image and video recognition tasks.

Main Methods:

• Proposed the Glance and Focus Network (GFNet), a sequential coarse-to-fine feature learning model.
• GFNet first extracts a global representation at low resolution, then focuses on salient regions for finer features.
• Utilized reinforcement learning to locate discriminant regions, requiring only classification labels.

Main Results:

• GFNet demonstrated remarkable efficiency across various image and video recognition tasks.
• Achieved a 1.3x reduction in average latency on an iPhone XS Max using MobileNet-V3 without accuracy loss.
• GFNet is compatible with various backbone models like MobileNets and EfficientNets.

Conclusions:

• GFNet offers an adaptive inference mechanism, terminating early when confident, thus avoiding redundant computations.
• The model significantly improves efficiency in visual recognition by addressing spatial redundancy.
• GFNet provides a flexible and efficient solution for image and video recognition challenges.