【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》-Toy模板网

这篇具有很好参考价值的文章主要介绍了【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》。希望对大家有所帮助。如果存在错误或未考虑完全的地方，请大家不吝赐教，您也可以点击"举报违法"按钮提交疑问。

【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms

ICCV-2017

1 Background and Motivation

NMS 是许多目标检测算法重要的步骤之一

【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms

NMS 的缺点，if an object lies within the predeﬁned overlap threshold, it
leads to a miss.（sets the score for neighboring detections to zero）

【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms
Intuitively, if a bounding box has a very high overlap with $M$ （maximum score）, it should be assigned a very low score, while if it has a low overlap, it can maintain its original detection score.

作者对 NMS 进行了改进

2 Related Work

NMS ﬁrst employed in edge detection techniques

3 Advantages / Contributions

NMS 基础上提出 Soft-NMS，not require any extra training and is simple to implement
obtain consistent improvements for the coco-style mAP metric on
standard datasets like PASCALVOC2007 (1.7% for both R-FCN and Faster-RCNN) and MS-COCO

4 Method

NMS 与 maximum score box 的 IoU 大于阈值的都抑制为了 0

【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms

全抑制为0，容易漏检，NMS should take the following conditions into account,

【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms
NMS 的阈值越高，越容易误检，越低约容易漏检

高 NMS 阈值情况下，mAP 反而会下降，因为 the increase in false positives would be much higher than the increase in true positives for this case because the number of objects is typically much smaller than the number of RoIs generated by a detector.

soft-NMS 大于阈值的分数降低

【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms
a list of detection boxes $B$ with scores $S$ .
maximum score $M$
the set of ﬁnal detections $D$

suppressing all nearby detection boxes with a low $N_t$ would increase the miss-rate.

其中 Rescoring Functions $f(iou(M, b_i))$ 的形式有如下两种

（1）非连续的（not continuous）
【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms
与最高分box 的 IoU 越大，分数降得更多（误检的概率比较大），IoU 越小，分数降的越少

IoU 等于 $N_t$ 的时候，值不一样

（2）连续的
【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms
计算复杂度同 traditional NMS

also a greedy algorithm and does not ﬁnd the globally optimal re-scoring of detection boxes.

不一定局限于指数形式，也可以是 Gompertz function

【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms

5 Experiments

5.1 Results

$N_t$ to 0.3， $\sigma$ to 0.5
【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms

【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms

Soft-NMS has more potential to improve recall at higher $O_t$ （detection evaluation threshold）.

【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms

5.2 Sensitivity Analysis

试试 soft-NMS 超参数 $\sigma$ 的敏感性

【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms

5.3 When does Soft-NMS work better?

（1）Localization Performance

千言万语，不如下面一个表来的直接
【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms
even when we obtain better performance at higher $O_t$ , performance at lower $O_t$ does not drop.

low values of $\sigma$ perform better at lower $O_t$ and higher values of sigma perform better at higher $O_t$ .

（2）Precision vs Recall

【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms

5.4 Qualitative Results

【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms

a large wide bounding box spanning multiple people is suppressed because it had a small overlap with multiple detection boxes with a higher score than it.

【Soft NMS】《Soft-NMS – Improving Object Detection With One Line of Code》,CNN,目标检测,目标跟踪,人工智能,soft nms