Enhancement of SSD by concatenating feature maps for object detection
BMVC 2017

本文是对SSD 的改进,通过牺牲一点速度来提高精度,主要解决SSD 两个问题:1)同一目标多次检测,2)小目标检测率不高
改进的地方:
1)将不同尺度的 特征图 融合起来;
2)增加 feature pyramid 网络层的特征图数量;
3)因为不同尺度特征图数量一样,可以对不同尺度特征图使用一个分类器

SSD 不同尺度的特征图是独立的,没有联系起来,导致同一目标在不同尺度上都被检测出来,进而导致同一目标多次检测

目标检测--Enhancement of SSD by concatenating feature maps for object detection

目标检测--Enhancement of SSD by concatenating feature maps for object detection

ConventionalSSDvs. theproposedRainbowSSD(R-SSD)
目标检测--Enhancement of SSD by concatenating feature maps for object detection

feature concatenation
目标检测--Enhancement of SSD by concatenating feature maps for object detection
这里介绍了三种特征图融合的方式:
1) pooling, 从前往后增加特征图,归一化尺寸
2) 借鉴图像分割中的 deconvolution, 从后往前增加特征图,归一化尺寸
3)本文的 Rainbow concatenation=pooling+deconvolution,这样每一个尺度的特征图数量都是一样的,这就导致后面可以使用一个分类器在不同尺度上检测
By using a single classifier, improvement on the generalization performance can be expected, and it can be effectively used for datasets with size
imbalance or for small datasets.

Increasing number of channels
目标检测--Enhancement of SSD by concatenating feature maps for object detection
The number of channels in each convolution layer are set to be 2 to 8 times larger than the original network

目标检测--Enhancement of SSD by concatenating feature maps for object detection

不同尺寸的目标检测精度
目标检测--Enhancement of SSD by concatenating feature maps for object detection

目标检测--Enhancement of SSD by concatenating feature maps for object detection