The use of object proposals is an effective recent approach for increasing the computational efficiency of object detection. We propose a novel method for generating object bounding box proposals using edges. Edges provide a sparse yet informative representation of an image. Our key insight is that the number of contours that are wholly enclosed within a bounding box is indicative of the likelihood of the box containing an object. We use this simple observation to score candidate boxes and generate proposals that rival the accuracy of Selective Search while being orders of magnitude faster.

Modern object detection methods such as R-CNN rely on object proposals — a small set of candidate regions that are likely to contain objects. The quality and speed of the proposal method directly impacts the overall detection pipeline. Selective Search, the most popular proposal method, is relatively slow, taking about 2 seconds per image. Our goal is to design a proposal method that is both accurate and fast enough for real-time applications.

Our method starts by computing an edge map using Structured Edge Detection. Given the edge map, we define an edge group as a set of connected edge pixels that share a common orientation. For each candidate bounding box, we compute a score based on the number of edge groups that are wholly contained within the box. Specifically, an edge group contributes to the score if its endpoints both lie within the box boundaries, indicating a closed contour.

A naive implementation would be slow due to the large number of candidate boxes. We achieve efficiency through several techniques: (1) using a sliding window with multiple scales and aspect ratios, (2) computing box scores incrementally using integral images over the edge map, and (3) applying non-maximum suppression to reduce redundant proposals. These optimizations allow Edge Boxes to generate proposals in 0.25 seconds per image, roughly 8x faster than Selective Search.

We evaluate on PASCAL VOC 2007 using recall at different numbers of proposals and IoU thresholds. With just 1,000 proposals, Edge Boxes achieves a recall of 0.81 at IoU 0.5, comparable to Selective Search's 0.82 with 2,000 proposals. At IoU 0.7, Edge Boxes with 1,000 proposals achieves 0.58 recall versus 0.52 for Selective Search, demonstrating better localization quality. When combined with R-CNN for detection, Edge Boxes achieves comparable mAP while being significantly faster.

We have proposed Edge Boxes, a simple, fast, and accurate method for generating object proposals from edges. Our key contribution is the insight that closed contours within a bounding box are strong indicators of objects. Edge Boxes provides an excellent trade-off between speed and accuracy, making it suitable as a drop-in replacement for slower proposal methods in modern detection pipelines.