16.Edge Boxes_ Locating Object Proposals from Edges.pdf

海尔

252

15页

0次

2021-02-23

50墨值下载

Edge Boxes: Locating

Object Proposals from Edges

C. Lawrence Zitnick and Piotr Doll´ar

Microsoft Research

Abstract. The use of object proposals is an eﬀective recent approach

for increasing the computational eﬃciency of object detection. We pro-

pose a novel method for generating object bounding box proposals us-

ing edges. Edges provide a sparse yet informative representation of an

image. Our main observation is that the number of contours that are

wholly contained in a bounding box is indicative of the likelihood of the

box containing an object. We propose a simple box objectness score that

measures the number of edges that exist in the box minus those that

are members of contours that overlap the box’s boundary. Using eﬃcient

data structures, millions of candidate boxes can be evaluated in a fraction

of a second, returning a ranked set of a few thousand top-scoring propos-

als. Using standard metrics, we show results that are signiﬁcantly more

accurate than the current state-of-the-art while being faster to compute.

In particular, given just 1000 proposals we achieve over 96% object recall

at overlap threshold of 0.5 and over 75% recall at the more challenging

overlap of 0.7. Our approach runs in 0.25 seconds and we additionally

demonstrate a near real-time variant with only minor loss in accuracy.

Keywords: object proposals, object detection, edge detection

1 Introduction

The goal of object detection is to determine whether an object exists in an

image, and if so where in the image it occurs. The dominant approach to this

problem over the past decade has been the sliding windows paradigm in which

object classiﬁcation is performed at every location and scale in an image [1–

3]. Recently, an alternative framework for object detection has been proposed.

Instead of searching for an object at every image location and scale, a set of

object bounding box proposals is ﬁrst generated with the goal of reducing the

set of positions that need to be further analyzed. The remarkable discovery

made by these approaches [4–11] is that object proposals may be accurately

generated in a manner that is agnostic to the type of object being detected.

Object proposal generators are currently used by several state-of-the-art object

detection algorithms [5, 12, 13], which include the winners of the 2013 ImageNet

detection challenge [14] and top methods on the PASCAL VOC dataset [15].

High recall and eﬃciency are critical properties of an object proposal gen-

erator. If a proposal is not generated in the vicinity of an object that object

2 C. Lawrence Zitnick and Piotr Doll´ar

Fig. 1. Illustrative examples showing from top to bottom (ﬁrst row) original image,

(second row) Structured Edges [16], (third row) edge groups, (fourth row) example

correct bounding box and edge labeling, and (ﬁfth row) example incorrect boxes and

edge labeling. Green edges are predicted to be part of the object in the box (w

) = 1),

while red edges are not (w

) = 0). Scoring a candidate box based solely on the

number of contours it wholly encloses creates a surprisingly eﬀective object proposal

measure. The edges in rows 3-5 are thresholded and widened to increase visibility.

cannot be detected. An eﬀective generator is able to obtain high recall using a

relatively modest number of candidate bounding boxes, typically ranging in the

hundreds to low thousands per image. The precision of a proposal generator is

less critical since the number of generated proposals is a small percentage of the

total candidates typically considered by sliding window approaches (which may

evaluate tens to hundreds of thousands of locations per object category). Since

object proposal generators are primarily used to reduce the computational cost

of the detector, they should be signiﬁcantly faster than the detector itself. There

is some speculation that the use of a small number of object proposals may even

improve detection accuracy due to reduction of spurious false positives [4].

In this paper we propose Edge Boxes, a novel approach to generating object

bounding box proposals directly from edges. Similar to segments, edges provide

a simpliﬁed but informative representation of an image. In fact, line drawings of

of 15

50墨值下载

database

关注

评论