[Paper note] PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection
2016-12-16 12:58
686 查看
paper
Author: Kye-Hyeon Kim, Sanghoon Hong, Byungseok Roh, Yeongjae Cheon, and Minje Park
Intel Imaging and Camera Technology
This network is smaller and more efficient than ResNet and can be a substitution of the latter.
In early stages, output nodes tend to be paired, i.e. one node’s activation is the opposite of another’s.
C.ReLU reduce the output channels by half and concatenate with negation.
Inception
Not yet been widely applied.
Cost-efficient building block for multi-scale representation.
1x1 Conv can preserve the receptive field of the previous layer.
Multi-scale representation like HyperNet
Concatenate the output of the last layer and two intermediate layers, whose size are 2x and 4x of the last layer.
Set 2x layer as reference scale, down-scaling (pooing) 4x layer, up-scaling (interpolation) the last layer.
Add Batch normalization layers before all ReLU.
Plateau detection based learning rate policy.
Measure the moving average of loss
Decide as on-plateau if its improvement is below a threshold.
Decrease the learning rate by a constant factor when on-plateau.
Number of proposals = 200
VOC2007 & VOC2012 performance
Author: Kye-Hyeon Kim, Sanghoon Hong, Byungseok Roh, Yeongjae Cheon, and Minje Park
Intel Imaging and Camera Technology
Highlight
Speed up (real-time) detection process with a more efficient feature extraction CNN, without lossing too much accuracy.This network is smaller and more efficient than ResNet and can be a substitution of the latter.
Main structure
C.ReLU: Concatenated ReLU in early stage of CNN to reduce the number of computation.In early stages, output nodes tend to be paired, i.e. one node’s activation is the opposite of another’s.
C.ReLU reduce the output channels by half and concatenate with negation.
Inception
Not yet been widely applied.
Cost-efficient building block for multi-scale representation.
1x1 Conv can preserve the receptive field of the previous layer.
Multi-scale representation like HyperNet
Concatenate the output of the last layer and two intermediate layers, whose size are 2x and 4x of the last layer.
Set 2x layer as reference scale, down-scaling (pooing) 4x layer, up-scaling (interpolation) the last layer.
Experiment
Training detailsAdd Batch normalization layers before all ReLU.
Plateau detection based learning rate policy.
Measure the moving average of loss
Decide as on-plateau if its improvement is below a threshold.
Decrease the learning rate by a constant factor when on-plateau.
Number of proposals = 200
VOC2007 & VOC2012 performance
相关文章推荐
- READING NOTE: PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection
- 多尺度R-CNN论文笔记(4): PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection
- PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection
- PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection
- PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection - arxiv 2016.08
- 目标检测--PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection
- 目标检测--PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection
- 读书笔记PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection
- PVANET----Deep but Lightweight Neural Networks for Real-time Object Detection论文记录
- 论文笔记:PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection
- 论文阅读--PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection
- PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection
- tensorfolw配置过程中遇到的一些问题及其解决过程的记录(配置SqueezeDet: Unified, Small, Low Power Fully Convolutional Neural Networks for Real-Time Object Detection for Autonomous Driving)
- 论文阅读:Deep Neural Networks for Object Detection
- 【深度学习论文笔记:Recognition】:Deep Neural Networks for Object Detection
- Deep Neural Networks for Object Detection论文翻译
- Deep Neural Networks for Object Detection(基于DNN的对象检测)
- 目标检测--Wide-Residual-Inception Networks for Real-time Object Detection
- 【Paper】DeepSaliency: Multi-Task Deep NeuralNetwork Model for Salient Object Detection
- 论文笔记 《Deep Neural Networks for Object Detection》