- 博客(122)
- 资源 (1)
- 收藏
- 关注

原创 秋招投递记录
牛客面试问题汇总(1)为什么LR模型损失函数使用交叉熵不用均方差?(2)YOLOv3如何确定正负anchor?YOLOv3不使用ohem表现良好,而SSD必须用ohem?(3)BN的前提假设?BN为什么可以防止过拟合?alpha和beta两个参数如何平移和缩放?BN先归一化再恢复是否矛盾。(4)L1、L2正则和weight delay这个参数的关系(5)sigmoid为什么...
2019-09-02 11:27:08
620

原创 【Python学习笔记】python小技巧
1、交换变量a,b = 1,2a,b = b,a2、将list中的所有元素转为单个字符串a =["I","am",&
2019-08-06 16:01:24
334

原创 【caffe学习笔记】caffe问题集
1、caffemodel和solverstateAny examples of using .solverstate with python interface2、scale和bias层Scale and Bias Layers Add ScalarLayer to multiply two Blobs with broadcasting
2019-06-13 16:26:41
150

原创 【目标检测】Faster R-CNN
相关知识:(部分图片来自目标检测之RCNN,SPP-NET,Fast-RCNN,Faster-RCNN)R-CNNR-CNN网络其实就是一个分类器,对提取到的候选区域(region proposal)进行分类,它的准确度主要依赖于候选区域算法。SSP流程:(1)在原始图像上用Selective Search算法得到候选区域(2)将整张图片前馈神经网络然后得到conv featur...
2018-11-15 17:32:40
2212
2
原创 nms实现
2.选出置信度最高的预测框,确认其为正确预测,并计算其与剩余预测框的IOU。3.若IOU>threshold阈值表示重叠度过高,将其删除。1.对所有预测框的置信度降序排序。4.重复上述过程处理完所有检测框。
2023-03-04 23:36:06
199
原创 Pytorch 打印网络中某层的参数
for name in net.state_dict(): print(name) print(net.state_dict()[name]) #. 参数for name, parameters in net.named_parameters(): print(name, ':', parameters)for (name, module) in net.named_modules(): print(name, ':', module)
2021-10-26 14:44:00
803
原创 PyTorch】常见错误: RuntimeError:one of the variables needed for gradient computation has been modified
找到网络模型中的 inplace 操作,将inplace=True改成 inplace=False;将网络结构中的+=操作进行修改,如下所示:out = out + res # not inplaceout += res # inplace
2021-06-17 18:38:02
1732
原创 c++联通域
https://www.pythonf.cn/read/124208https://www.pythonf.cn/read/124208
2021-03-15 19:17:46
247
原创 小目标检测算法调研
1、Scale Match for Tiny Person Detection论文链接:https://arxiv.org/abs/1912.10664代码链接:https://github.com/ucas-vg/TinyBenchmark专为远距离和大背景下的微小目标设计的检测器2、RetinaNet升级版:TinaFace论文链接:https://arxiv.org/abs/2011.13183代码链接:https://github.com/Media-Smart/vedadet/tree
2020-12-18 15:44:42
979
1
原创 【语义分割】Self-Correction for Human Parsing
The ambiguous boundary between different semanticparts and those categories with similar appearance usuallyare confusing, leading to unexpected noises in ground truthmasks. To tackle the problem of learning with label noises,this work introduces a puri
2020-12-15 15:18:02
1474
原创 andriod和ios抓取log
andriod:adb logcat > log.txtiosidevicesyslog >> iphone.log &
2020-09-07 17:05:23
664
原创 【基础知识】blender环境配置
参考:1、blender python2、 blender configuration - 缺少某些包3、 Install external package in Blender (python websockets)
2020-06-22 22:11:38
3264
原创 【姿态估计】相机内参外参
基础知识:世界坐标系、相机坐标系、图像坐标系、像素坐标系之间的关系相机内参:相机坐标到像素坐标相机外参:世界坐标到相机坐标
2020-06-18 14:07:34
1627
原创 item2卡顿
https://n3xtchen.github.io/n3xtchen/osx/2017/01/18/osx-iterm2vim-slowhttps://www.v2ex.com/t/238097https://www.logcg.com/archives/2376.htmlhttps://answer-id.com/70897309http://www.classnotes.cn/1632.html
2020-05-31 23:14:47
618
原创 TypeError: Expected cv::UMat for argument 'img'
img = img.astype(np.uint8)
2020-05-08 13:46:37
14421
15
原创 【基础知识】tensorboardX问题集合
1、‘GraphDef’ object does not support indexingtensorboardX需要更新pip install --upgrade tensorboardX2、RuntimeError: Only tensors or tuples of tensors can be output from traced functions解决方案...
2020-04-26 14:44:45
1660
原创 【目标检测】目标检测评估指标mAP
1、将所有预测框根据置信度从大到小进行排序;2、选取不同的置信度阈值,预测框的置信度大于此阈值为Positive,小于为negative;3、确定IOU阈值(一般0.5)。计算Positive与每个GT的IOU,当最大的IOU大于0.5时,为TP;小于0.5时为FP;FP 误检;FN漏检precision=TPTP+FP\text {precision}=\frac{T P}{T P+...
2020-04-24 17:44:53
1349
1
原创 【Pytorch】Pytorch Dataloader加速
1、https://discuss.pytorch.org/t/how-to-speed-up-the-data-loader/13740/32、h5py database3、使用h5py库读写超过内存的大数据
2019-11-03 14:58:43
555
原创 【深度学习】知识总结
1、语义分割和目标检测具有相似之处分辨率和感受野的关系下采样会损失细节信息,但减少下采样倍数,会减小感受野,使用空洞卷积。空洞卷积设计不慎会产生网格效应。车道线分割中的自蒸馏方法SAD,可使得特征层学习到更丰富的全局上下文信息浅层信息和深层信息的融合:skip connection 、FPN语义分割中有很多模块可实现特征的融合:如CAB和FFM,采用channel attention机制...
2019-09-27 12:10:27
493
原创 【深度学习】Focal Loss
FL(pt)=−αt(1−pt)γlog(pt)\mathrm{FL}\left(p_{\mathrm{t}}\right)=-\alpha_{\mathrm{t}}\left(1-p_{\mathrm{t}}\right)^{\gamma} \log \left(p_{\mathrm{t}}\right)FL(pt)=−αt(1−pt)γlog(pt)pt={p if&...
2019-09-20 18:51:09
327
原创 【Portrait分割】BANet:Boundary-Aware Network for Fast and High-Accuracy Portrait Segmentation
文章目录Abstract1、Background2、Method of BANet2.1、对gt做Canny检测可参考:DFNAbstract与通用语义分割任务相比,Portrait分割需要更高的精度和更快的速度。Boundary-Aware Network (BANet) 选择性提取边界区域的细节信息获得高质量的分割结果,且可做到实时分割( ≥ 25FPS)同时BANet设计了一个r...
2019-08-16 19:10:24
1685
原创 【轻量级网络】MobileNetV2_ Inverted Residuals and Linear Bottlenecks
paper:MobileNetV2code:caffe pytorch文章目录1、Preliminaries, discussion and intuition1.1、Depthwise Separable Convolutions1.2、Linear BottlenecksMobilenetv2核心模块是:inverted residual structure.1、Prelimin...
2019-08-16 13:58:12
364
原创 【轻量级网络】MobileNetV1: Efficient Convolutional Neural Networks for Mobile Vision Applications
paper:MobileNetV2code:caffe pytorch文章目录1、Preliminaries, discussion and intuition1.1、Depthwise Separable Convolutions1.2、Linear BottlenecksMobilenetv2核心模块是:inverted residual structure.1、Prelimin...
2019-08-16 11:00:39
243
原创 【语义分割】context module
ASPP - Atrous Spatial Pyramid Poolingtensorflow实现:networkASPP是deeplabv3中的核心模块。ASPP采用三个并行的空洞卷积,空洞卷积rates = (6, 12, 18),从而提取到多尺度的特征。另外,为了更好融合全局上下文信息,ASPP还采用了全局平均池化(池化+BN+bilinearly upsample)PPM -...
2019-08-15 14:11:35
1976
原创 【Freespace分割】Real-Time Freespace Segmentation on Autonomous Robots for Detection of Obstacles
paper:Real-Time Freespace Segmentation
2019-08-15 11:59:14
1440
1
原创 【Prtrait分割】ExtremeC3Net: Extreme Lightweight Portrait Segmentation Networks using C3-modules
paper :ExtremeC3Net文章目录ExtremeC3Net is a lightweight and robust portrait segmentation algorithm consisting of a two-branched architecture based on the concentratedcomprehensive convolutions block....
2019-08-14 18:18:57
2112
1
原创 【车道线检测】Multi-Class Lane Semantic Segmentation using Efficient Convolutional Networks
paper:Multi-Class Lane Semantic Segmentation using Efficient Convolutional NetworksAbstract车道线检测算法流程:lane feature extraction,feature segment grouping, and lane model fitting.这篇论文提出了两个方法:Feature Si...
2019-08-13 17:19:46
683
原创 【车道线检测】SAD -- Learning Lightweight Lane Detection CNNs by Self Attention Distillation
paper:Learning Lightweight Lane Detection CNNs by Self Attention DistillationCode:githubAbstractWithout learning from much richer context, these models often fail in challenging scenarios, e.g. ,s...
2019-08-13 14:04:17
50404
6
原创 【语义分割】ENet -- A Deep Neural Network Architecture for Real-Time Semantic Segmentation
Network architectureENet的intial下采样使用的是结构(a),后续的下采样均使用结构(b):ENet的通用单元结构如下:第一个1x1降维,第二个1x1升维ENet不使用BNDesign choices1、Feature map resolution...
2019-08-07 17:13:07
421
原创 【深度学习】Understanding the Effective Receptive Field in Deep Convolutional Neural Networks
paper:Understanding the Effective Receptive Field in Deep Convolutional Neural NetworksAn effective receptive field both has a Gaussian distribution and only occupies a fraction of the full theoreti...
2019-08-06 22:58:13
1155
空空如也
TA创建的收藏夹 TA关注的收藏夹
TA关注的人