快速小目标检测--Feature-Fused SSD: Fast Detection for Small Objects
2017-09-19 16:22
507 查看
Feature-Fused SSD: Fast Detection for Small Objects
本文针对小目标检测问题,对 SSD 模型进行了一个小的改进,将 contextual information 引入到 SSD 中 帮助SSD检测小目标。
contextual information 对于小目标的检测 重要性是不言而喻的。小目标在图像中 limited resolution and information,只能借助其周边信息了。
Feature-fused SSD architecture
这里主要还是讲不通尺度的特征图信息融合起来。
Which layers to combine? 但是具体融合哪个卷积层的特征图了
Because of SSD predicting small objects with its shallower layers,所以我们不用对应大目标的 deeper layers
For choosing the proper feature fusion layers, effective receptive fields in different layers are explored with deconvolution method
对应上图中的小船,SSD中的 conv4_3 对应的 effective receptive field 是最合适的。 其他层的特征都不是很好
shallower layers 具有 contextual information ,但是它们没有 sematic information,deeper layers 具有 sematic information,所以我们需要将这两类信息融合起来。这里我们尝试了两种融合策略。
concatenation module and element-sum module
3.2 Concatenation Module
3.3 Element-Sum Module
4 Experimental Resultis
Detection results of different fusion layers
Detection results of different number of kernels
Results on PASCAL VOC2007 test set (with IOU=0.5)
The running time illustration of different models
本文针对小目标检测问题,对 SSD 模型进行了一个小的改进,将 contextual information 引入到 SSD 中 帮助SSD检测小目标。
contextual information 对于小目标的检测 重要性是不言而喻的。小目标在图像中 limited resolution and information,只能借助其周边信息了。
Feature-fused SSD architecture
这里主要还是讲不通尺度的特征图信息融合起来。
Which layers to combine? 但是具体融合哪个卷积层的特征图了
Because of SSD predicting small objects with its shallower layers,所以我们不用对应大目标的 deeper layers
For choosing the proper feature fusion layers, effective receptive fields in different layers are explored with deconvolution method
对应上图中的小船,SSD中的 conv4_3 对应的 effective receptive field 是最合适的。 其他层的特征都不是很好
shallower layers 具有 contextual information ,但是它们没有 sematic information,deeper layers 具有 sematic information,所以我们需要将这两类信息融合起来。这里我们尝试了两种融合策略。
concatenation module and element-sum module
3.2 Concatenation Module
3.3 Element-Sum Module
4 Experimental Resultis
Detection results of different fusion layers
Detection results of different number of kernels
Results on PASCAL VOC2007 test set (with IOU=0.5)
The running time illustration of different models
相关文章推荐
- 目标检测--Enhancement of SSD by concatenating feature maps for object detection
- 目标检测--A Unified Multi-scale Deep Convolutional Neural Network for Fast Object Detection
- 目标检测“Perceptual Generative Adversarial Networks for Small Object Detection”
- 目标检测“A-Fast-RCNN: Hard Positive Generation via Adversary for Object Detection”
- 目标检测“Feature Pyramid Networks for Object Detection”
- 对抗学习用于目标检测--A-Fast-RCNN: Hard Positive Generation via Adversary for Object Detection
- 目标检测--Feature Pyramid Networks for Object Detection
- 【深度学习:目标检测】RCNN学习笔记(1):Rich feature hierarchies for accurate object detection and semantic segmentat
- 【深度学习:目标检测】RCNN学习笔记(2):Rich feature hierarchies for accurate object detection and semantic segmentat
- 对抗学习用于目标检测--A-Fast-RCNN: Hard Positive Generation via Adversary for Object Detection
- 目标检测“Perceptual Generative Adversarial Networks for Small Object Detection”
- 行人检测论文笔记:Fast Feature Pyramids for Object Detection?
- 特征金字塔特征用于目标检测:Feature Pyramid Networks for Object Detection
- 视频目标检测--Flow-Guided Feature Aggregation for Video Object Detection
- 嵌入式目标检测--Fast YOLO: A Fast You Only Look Once System for Real-time Embedded Object Detection
- 目标检测 Feature Pyramid Networks for Object Detection(FPN)论文笔记
- 对抗学习用于目标检测--A-Fast-RCNN: Hard Positive Generation via Adversary for Object Detection
- 行人检测论文笔记:Fast Feature Pyramids for Object Detection?
- 目标检测“Perceptual Generative Adversarial Networks for Small Object Detection”
- 快速多尺度人脸检测--Multi-Scale Fully Convolutional Network for Fast Face Detection