Scale Normalized Image Pyramids With AutoFocus for Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Object detection is crucial for computer vision tasks.
• Existing methods often face computational challenges, especially with large datasets and complex scenes.
• Human vision's foveal nature, focusing on relevant areas, offers a potential model for efficient processing.

Purpose of the Study:

• To introduce an efficient foveal framework for object detection.
• To improve training speed and inference efficiency in object detection models.
• To develop a method that mimics human visual attention for computational gains.

Main Methods:

• Generation of a scale-normalized image pyramid (SNIP) to focus on objects within specific size ranges at different scales.
• Implementation of an efficient spatial sub-sampling scheme (SNIPER) during training, utilizing known object locations to reduce computational cost.
• Adoption of a coarse-to-fine approach (AutoFocus) during inference, predicting object-like regions for successive processing across image pyramid scales.

Main Results:

• SNIPER achieves up to a 3× speed-up during training compared to standard methods.
• AutoFocus, when combined with SNIP, results in a 2.5–5× speed-up during inference.
• The proposed methods enhance object detection accuracy by improving the learning of object-sensitive filters.

Conclusions:

• The SNIPER framework offers significant training efficiency for object detection.
• AutoFocus effectively accelerates inference by adaptively focusing on relevant image regions.
• This foveal approach provides a computationally efficient and accurate solution for object detection tasks.