![](https://img-blog.csdnimg.cn/20201014180756913.png?x-oss-process=image/resize,m_fixed,h_64,w_64)
paper reading
DreamLike_zzg
这个作者很懒,什么都没留下…
展开
-
salient object detection
A Simple Pooling-Based Design for Real-Time Salient Object Detection (2019 CVPR)Global Guidance Module (GGM) : PPM feature 与 FPN 每个level 融合。增加RF,缓解obj预测不完整。Feature Aggregation Module (FAM): 类似PPM,add代替concate。 1、reducing the aliasing effect of ups...原创 2020-06-04 18:18:00 · 624 阅读 · 0 评论 -
Knowledge Distillation for Segmentation 笔记
目录Inter-Region Affinity Distillation for Road Marking Segmentation (2020.04)Learning Lightweight Lane Detection CNNs by Self Attention Distillation (2019.08)Knowledge Adaptation for Efficient Semantic Segmentation(CVPR 2019)Inter-Region Affinity Di原创 2020-05-22 17:15:04 · 717 阅读 · 1 评论 -
Knowledge Distillation 笔记
Inter-Region Affinity Distillation for Road Marking Segmentation (2020.04)用 inter-region affinity graph 描述structural relationship。每个node表示一种class的 areas of interest (AOI)(同类算一个 or instance算一个?),...原创 2020-04-15 16:33:49 · 1430 阅读 · 0 评论 -
轻量化backbone
MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications (2017.04)1、使用 depthwiseconv + pointwise conv代替 standard conv,计算量减少比例:计算量和参数量主要集中在 pointwise conv上,pointwise...原创 2020-04-09 11:56:53 · 929 阅读 · 0 评论 -
one/few-shot segmentation 笔记
数据集pascal-5i每个sub-dataset中,trainset来自pascal trainset,testset来自pascal valsetOne-Shot Learning for Semantic Segmentation (Sep 2017)使用dataset:pascal-5itrain时,support 和 query 均从Dtrain中抽样,cla...原创 2020-03-18 17:33:07 · 2151 阅读 · 3 评论 -
DUC HDC 笔记
Understanding Convolution for Semantic Segmentation1 Jun 2018thought1、增加的可学习参数形式上不一定对应,能变换回来就行2、直觉上的不合理要仔细分析motivation1、上采样更好地恢复信息,bilinear upsampling is not learnable, deconvolution zeros have...原创 2020-02-02 20:41:39 · 3481 阅读 · 0 评论 -
deeplab笔记
deeplab V1backbone:VGG-16 + CRF最后两层max pooling (pool4, pool5) 的 stride = 1最后三个conv层(conv5_1、conv5_2、conv5_3)的dilate rate设置为2,第一个全连接层的dilate rate设置为4during trainingtarget 为 GT label subsampled by...原创 2020-02-01 19:46:10 · 205 阅读 · 0 评论 -
RefineNet笔记
RefineNet: Multi-Path Refinement Networks for High-Resolution Semantic Segmentationhttps://arxiv.org/pdf/1611.06612.pdfThought1、多利用residual 结构, 包括广义的,利于优化2、与 FPN 同时间出现,当时流行特征融合institution1、subsa...原创 2020-01-30 13:54:23 · 311 阅读 · 0 评论 -
FDNet笔记
Learning Fully Dense Neural Networks for Image Semantic Segmentation1、融合feature map时,concate所有前层feature map,encoder network is based on the DenseNet-2642、boundary-aware loss:边界像素作为hard exampleArchi...原创 2020-01-29 18:26:27 · 1080 阅读 · 0 评论 -
U-Net笔记
U-Net: Convolutional Networks for Biomedical Image Segmentation18 May 2015Network ArchitectureThe architecture consists of a contracting path to capture context and a symmetric expanding path that ...原创 2020-01-29 17:32:23 · 245 阅读 · 0 评论 -
Understanding the Effective Receptive Field in Deep Convolutional Neural Networks
相关参考https://zhuanlan.zhihu.com/p/86599013ERF比理论感受野小近似成高斯分布subsampling、diation conv 可增大ERF训练过程中ERF不断增大原创 2019-11-20 13:39:19 · 342 阅读 · 0 评论 -
M2Det 阅读
@InProceedings{author = {Qijie Zhao, Tao Sheng1,Yongtao Wang, Zhi Tang, Ying Chen, Ling Cai and Haibin Ling},title = {M2Det: A Single-Shot Object Detector based on Multi-Level Feature Pyramid Networ...原创 2019-04-28 16:08:11 · 852 阅读 · 2 评论 -
FSAF 阅读
@InProceedings{author = {Chenchen Zhu, Yihui He, Marios Savvides},title = {Feature Selective Anchor-Free Module for Single-Shot Object Detection},booktitle = {The IEEE Conference on Computer Vision...原创 2019-04-22 21:24:09 · 2030 阅读 · 2 评论 -
Focal Loss 阅读
论文:Focal Loss for Dense Object Detection论文链接:https://arxiv.org/abs/1708.02002来龙去脉分析:https://blog.csdn.net/c9Yv2cf9I06K2A9E/article/details/78920998Loss 梯度推导:https://github.com/zimenglan-sysu-512/p...原创 2019-04-11 13:10:30 · 169 阅读 · 0 评论 -
RefineDet 阅读
@InProceedings{author = { Shifeng Zhang, Longyin Wen, Xiao Bian, Zhen Lei1, Stan Z. Li },title = { Single-Shot Refinement Neural Network for Object Detection },booktitle = {The IEEE Conference on C...原创 2019-04-06 11:09:56 · 369 阅读 · 1 评论