luoyizhi516's Projects
定位综述 目前用于人脸定位的有两种主流方法,一种是基于级联形状回归模型,另一种就是基于深度学习的方法。级联形状回归方法就是使用回归模型,直接学习从人脸表征到人脸形状的映射函数,进而建立从表观到形状的对应关系。这种方法的学习依赖于训练集的选取,如果训练集中包含了复杂的姿态变化,学习到的函数测试性能就会比较好。现有很多基于回归的方法,其中比较突出的有颜水成的DCR(Deep Cascaded Regression)、ESR(Explicitly Shape Regression)方法、LBF(Local Binary Features)方法以及SDM(Supervised Decent Method)方法。但是,上述方法也存在一定的问题,例如,基于提取到的局部图像特征来定位时很难找到具有全局信息的关键点。基于点分布模型的方法和基于形状回归的方法都很依赖于初始值,其中,初始值通常由训练集的平均形状来给定,如果初始值或者设定的平均形状远远偏离目标位置,很难收敛到正确位置。比如,训练集中大部分都是正脸,那么对于测试集中大偏转角度人脸图像的定位就比较困难。 于是引入深度学习来解决较大面部偏转姿态的回归问题,深度学习的最大的优点是有强大的表达能力,可以自学习图像的特征,不需要人为的寻找特征。目前用的基本网络有(1)级联方式:由粗到精一级一级的优化前一步得到的形状,比如香港中文大学汤晓鸥老师的研究团队采用深度卷积神经网络的3个级联结构,逐步细化特征点位置。(2)沙漏网络:密集堆叠连接的U-Nets来进行人脸关键点定位,如CU-Nets通过卷积-反卷积,下采样与上采样网络,跨越不同U-Nets进行全局梯度传播,融合多尺度特征,不断迭代优化关键点坐标。但是深度学习方法的缺点是模型过于复杂,参数非常多,耗时比较长。 深度学习的国内外方法优化:(1)最早的有将人脸进行五官分区域定位回归,但很容易使最终结果陷入局部最优值而不是全局最优。(2)heatmap热图的方法,对数据标签进行处理生成高斯热图,从而更好的回归关键点位置。(3)风格聚合方法,对数据集进行处理生成不同风格的图片,以应对复杂的真实环境如光照强弱。(4)基于边缘感知的人脸关键点检测算法,首先通过消息传递并结合对抗学习得到高精度的边缘线检测结果,再将边缘线信息融合到关键点检测中,来提升算法在大侧脸、夸张表情、遮挡、模糊等极端情况下的鲁棒性。(5)3D人脸模型方法,一种方式是结合3D人脸姿态估计与投影来确定特征点初始位置,然后使用经典的回归树集成ERT方法来更好的进行位置回归;另一种是直接从单幅人脸图像中同时回归出3D人脸结构和密集对齐点;该类方法通过回归位置图,来获得3D几何以及语义信息。(6)人脸特征点检测与形状拟合方法,基于“特征提取”+“回归坐标”的方法,把形状拟合的坐标回归问题,转化为坐标PCA压缩后系数与形状整体仿射变换系数的回归问题(相当于将传统的SDM算法用于深度学习算法当中)。 目前定位方法的改进是: 1.基于深度学习的方法:改进如下: (1)受人体关键点定位 难例挖掘方法的启发,在全局的globalnet之后继续对损失较大的点继续进行微调。 (2)并行集成的方式同时训练两个网络globalnet、refinenet,测试时,第一个网络输出的l2 loss 取top k loss作为refinenet的辅助损失继续微调回归关键点位置。(在300w 有缺失块的数据集上最终的测试结果,平均定位误差为5.9%) 300-W数据库: 68点定位库,包含4个数据集,比如AFW,LFPW,HELLE,还包含135张IBUG里面的图片,IBUG图片对于定位来说都是非常具有挑战性的,不管是旋转角度还是遮挡程度都是非常大的。我们用HELEN的2000张,LFPW的811张,AFW的337张,总共3148张图片作为训练集。用HELEN的554张,IBUG的135张图片,总共689张图片作为测试集。一般测试集分为3个标准:简单,有挑战,全集。简单的测试集就是前面554张图片,有挑战的就是后面那135张图片,全集即689张图片。对三个测试集分别测试,得出测试结果,并与其它方法进行比较。测试标准为定位的平均误差,是被双瞳孔距离归一化后的,平均误差越低代表这一算法越好。
Two-Stream Adaptive Graph Convolutional Networks for Skeleton-Based Action Recognition in CVPR19
Implementation of Action Recognition using 3D Convnet on UCF-101 dataset.
This is the full analysis of the forward, inverse kinematics, trajectory planning, path planning, and controlling the end effector.
An accident avoidance program that raises alert when nearby vehicles are moving at a relative speed faster than a threshold value, additionally it logs some data onto NEM-Mijin blockchain network
Action recognition tutorial using UCF-101 dataset.
Some scripts used for action recognition on UCF101 dataset
recognices actions from the videos
Age estimation project in TDT4173, NTNU Trondheim. We have created a deep CNN for predicting ages based on images.
农业知识图谱(AgriKG):农业领域的信息检索,命名实体识别,关系抽取,智能问答,辅助决策
Hurricane Trajectory using Recurrent Neural Networks. In our model, the inputs to the Recurrent Neural Network (RNN) are wind speed, pressure, and aggregated features of direction, distance, and grid identification. The distance and direction values were calculated from the latitude and longitude points given by the Unisys dataset. Our RNN is learning from all these features and it is predicting which grid location the hurricane will be moving to next.
The people counter application demonstrates how to create a smart video IoT solution using Intel® hardware and software tools. The app will detect people in a designated area, providing the number of people in the frame, average duration of people in frame, and total count.
Github repository for the paper End-to-end Animal Image Matting
Awesome papers for markerless animal motion capture and 3D reconstruction.
applying optical flow and yolo detectors to detect animal (mostly primates) based on video
A database collection of farm animal behaviour phenotypes for modelling exploration.
An annotation and instance segmentation-based multiple animal tracking and behavior analysis package.
Python GUI: simple video annotator for animal behavior
Multi class audio classification using Deep Learning (MLP, CNN): The objective of this project is to build a multi class classifier to identify sound of a bee, cricket or noise.
大家好,这是cv调包侠开源原创项目,您可以特别方便地通过我的auto-Maker实现目标检测数据集的实时生成,包括:实时数据采集、自动标注、转换、增强,并可以直接进行yolov3、yolov4、yolov5、eficientdet等,它可以直接导出到onnx,通过openvino和tensor-RT加速,除了检测外,还支持分类算法,可以在一分钟内完成图像的智能分类。欢迎来star Hello everyone, this is me: cver open source original project, you can particularly convenient through my auto_ Maker realizes the real-time production of target detection data set, including: real data acquisition, automatic annotation, conversion, enhancement, and can directly carry out yolov3, yolov4 , yolov5, eficientdet, etc., which can be directly exported to onnx and accelerated by openvino and tensor RT. besides detection, it also supports classification algorithm, which can complete intelligent classification of images in one minute. Welcome to star~
Abstract With the advancement of Deep Neural Networks (DNN), the accuracy of sound classification such as Urban Sound Classification, Environmental Sound Classification etc., has been significantly improved. In this project, we propose a model that uses Convolutional Neural Networks (CNN) to identify sound based on the spectrograms for different sound samples collected. The model can be used for detection of deforestation, detection of shooting in urban areas and detection of strange noises at odd hours in streets such as Air Conditioner, Car Horn, Children Playing, Dog bark, Drilling, Engine Idling, Gun Shot, Jackhammer, Siren, Street Music etc., Challenges Environmental sound work has two major obstacles, namely the lack of audio data labelled. Previous work focused on audio from carefully produced films or TV tracks from particular environments such as elevators or office spaces and commercial or proprietary datasets. Lack of fundamental vocabulary in Environmental Sounds work. This means that the classification of sounds in to the semantic groups may vary from study to study, making it difficult to compare results so the goal of this notebook is to address the two challenges mentioned above. Dataset The dataset is called UrbanSound8K and contains 8732 labelled sound excerpts (<=4s) of urban sounds from 10 classes: - The dataset contains 8732 sound excerpts (<=4s) of urban sounds from 10 classes, namely: Air Conditioner Car Horn Children Playing Dog bark Drilling Engine Idling Gun Shot Jackhammer Siren Street Music The attributes of data are as follows: ID Unique ID of sound excerpt Class type of sound Problem statement It will show how to apply Deep Learning techniques to environmental recognition sounds, focusing specifically on recognizing unique Environmental sounds. If we give an audio sample of a few seconds duration in a computer-readable format (such as a.wav file), we want to be able to determine whether it contains one of the target Environmental sounds with a corresponding classification accuracy score. Note: Loading audio files and pre-processing takes some times to complete with large dataset. To avoid reload every time reset the kernel or resume works on next day, all loaded audio data will be serialized into a object file. so next round only need to load the seriazed object file. Optional GPU configuration initialization
A machine learning approach of automatic identification and counting of all types of blood cells: RBCs, WBCs, and Platelets with K-nearest neighbor (KNN) and intersection over union (IOU) based verification.
draw ROC,PR curve and calculate AUC MAP IoU for image semantic segmentation problem
Open source technology for agriculture, farming, and gardening
:punch: 计算机视觉中用到的注意力模块和其他即插即用模块PyTorch Implementation Collection of Attention Module and Plug&Play Module
Awesome Crowd Counting
A selection of state-of-the-art research materials on trajectory prediction
A survey of occlusion handling in object detection
A curated list of resources dedicated to optical flow algorithms. Feel free to make PRs to contribute.