Any-scale object detection improved by scale-aware super-resolution images
Any-scale Object Detection using Arbitrary-scaled Images
Computer Vision and Pattern Recognition
Summary
Detecting objects in images often requires looking at different sizes, but current methods only check a few fixed sizes. The authors propose a way to zoom in or out on objects at any size using super-resolution images, which makes the detection more flexible. However, just zooming in too much can cause many false alarms, so they added a step to decide which zoom levels make sense for different parts of the image. This approach helps detect objects more accurately across a continuous range of sizes.
What this means in practice
- •For autonomous vehicle developers: Improve detection of objects at varying distances and sizes by using adaptive image scaling to reduce false positives in real-time systems.
- •For industrial inspection teams: Enhance detection accuracy for objects of varying sizes within industrial images by integrating arbitrary scale super-resolution guided by scale proposals.
Authors
Kazutoshi Akita, Norimichi Ukita
Abstract
This paper proposes any-scale object detection using arbitrary-scale super-resolution for continuously rescaling object images, while general multi-scale object detection uses discretely rescaled appearance representations. However, a naive usage of super-resolution produces many false-positive detections if many super-resolution images are independently fed into an object detector. Our method suppresses these false positives by predicting scale proposal maps, each of which represents a set of pixels appropriate for each super-resolution scale.