目标检测论文: Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample

Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection
PDF: https://arxiv.org/pdf/1912.02424.pdf
PyTorch: https://github.com/shanglianlm0525/PyTorch-Networks

1 概述

FCOS证明了anchor free的方法是work的,但是没解释为什么高了retinanet几个点。这篇文章从三个方面用大量实验来分析,得出FCOS涨点只和正负样本的采样方式有关, 和使用哪种box回归方式无关以及多个anchor采样策略时也没有关系.

RetinaNet:
目标检测论文: Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample_第1张图片
FCOS:
目标检测论文: Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample_第2张图片

2 Difference Analysis(差异分析)

FCOS相比RetinaNet有一些improvements,包括:

  • GroupNorm
  • GIoU
  • GT Box
  • Centerness
  • Scalar
    把这些improvements用在RetinaNet上, 发现FCOS和RetinaNet还有一些差距.
    目标检测论文: Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample_第3张图片

2-1 the positive/negative sample definition (如何界定正负样本)

  • RetinaNet通过计算anchor和groundtruth的IoU来完成的,这样做的缺陷很明显,anchor是正样本还是负样本非常受限于anchor的设计,如下图第一行;
  • FCOS通过判定feature上的每个点是否落入到groundtruth中来判定正负样本的,很显然这样就摆脱了类似于anchor这种受限于hard-craft的缺陷.
    目标检测论文: Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample_第4张图片
    实验表明正负样本选择策略对mAP有不小的影响.
    目标检测论文: Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample_第5张图片

2-2 the regression starting status (Regression的方式不同)

  • RetinaNet从预置anchor框中回归目标;
  • FCOS从点定位目标;
    目标检测论文: Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample_第6张图片
    从Table 2 中可以看出, 两者回归方式对性能影响很小.

2-3 the number of anchors tiled per location (anchor数量不同)

在传统的基于IOU选择正负样本策略下, 增加anchor数量可以明显增加mAP, 但是当选择正负样本策略更优时, anchor数量对mAP影响很小.
目标检测论文: Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample_第7张图片

3 Adaptive Training Sample Selection(ATSS)

ATSS过程很容易理解,替代直接使用IOU作为候选框的阈值,首先选择ground-truth中心最近的k个候选框,计算均值和方差作为IOU阈值,去除中心点在检测目标之外的候选框。整个过程中只有一个超参数k,且对模型影响小。
目标检测论文: Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample_第8张图片

你可能感兴趣的:(Deep,Learning,Paper,Reading)