目标检测的两大范式—‘top-down’ 和‘bottom-up’

两大范式可分为：‘top-down’ 和‘bottom-up’

引自《CenterNet++ for Object Detection》 RELATED WORK部分

DOI：10.1109/TPAMI.2023.3342120

“Bottom-up object detection by grouping extreme and center points”和“Object detection in 20 years: A survey” 两篇文献认为‘top-down’ 和‘bottom-up’是检测算法如何定位目标的方式。

1. Top-down

‘top-down’方法在预定义的关键点和网格之外，放置一系列anchors box或anchor point来定位目标。首先寻找代表整个目标的proposals，然后通过分类和回归proposals来确定目标类别和边界框。

根据不同形式的proposals，‘top-down’方法又可被进一步分为anchor-based和anchor-free。代表算法有：Faster RCNN，YOLO。

以YOLO为例，YOLO在特征图上每个点输出的回归框参数需与anchor box进行IoU匹配，从而确定正负样本。随后，再与ground truth计算loss。

2. Bottom-up

‘bottom-up’检测目标的单独部分，而不是将物体作为一个整体。随后，通过可训练的后处理算法聚类属于同一类别的单独部分。代表算法有：DPM，CornerNet，ExtremeNet。

以CenterNet为例，CenterNet直接在特征图上寻找唯一的峰值点作为中心点（x，y）（以heat map展示），回归高和宽的偏离值，随后与ground truth计算loss。

本文为笔记，纯属个人理解。如有不当之处，请指正。

3. 原文如下：

Here, we follow the convention [69], [72] in using ‘top-down’ and ‘bottom-up’ to indicate how the detection algorithm locate objects – the ‘top-down’ approaches locate objects by placing a number of anchors beyond pre-defined keypoints or grids, and the ‘bottom-up’ approaches instead organize the keypoints into objects.

Top-down approaches first find the proposals that represent the whole objects, and then determine the classes and the bounding boxes of the objects by classifying and regressing the proposals. The proposals could be further divided into anchor-based and anchor-free according to the different forms of proposals.

Bottom-up approaches detect the individual parts of objects instead of perceiving the objects as a whole. Subsequently, the individual parts that belong to the same object are grouped together by some trainable postprocessing algorithms.