1. 程式人生 > >人體姿態估計2014-2017

人體姿態估計2014-2017

一、常用資料集Pose Estimation/keypoint常用資料集1. Posetrack:https://posetrack.net/> 500 video sequences> 20K frames> 150K body pose annotations3 challenges2. LSP:http://sam.johnson.io/research/lsp.html樣本數:2K關節點個數:14全身,單人3. FLIC:https://bensapp.github.io/flic-dataset.html樣本數:2W關節點個數:9全身,單人4. MPII:http://human-pose.mpi-inf.mpg.de/樣本數:25K關節點個數:16全身,單人/多人,40K people,410 human activities5. MSCOCO:http://cocodataset.org/#download樣本數:>= 30W關節點個數:18全身,多人,keypoints on 10W people6. AI Challenge:https://challenger.ai/competition/keypoint/subject樣本數:21W Training, 3W Validation, 3W Testing關節點個數:14全身,多人,38W people二、主流方法2D Pose estimation主要面臨的困難:遮擋、複雜背景、光照、真實世界的複雜姿態、人的尺度不一、拍攝角度不固定等。單人姿態估計傳統方法:基於Pictorial Structures, DPM▪ 基於深度學習的演算法包括直接回歸座標(Deep Pose)和通過熱力圖迴歸座標(CPM, Hourlgass)目前單人姿態估計,主流演算法是基於Hourlgass各種更改結構的演算法。多人姿態估計二維影象姿態估計基於CNN的多人姿態估計方法,通常有2個思路(Bottom-Up Approaches和Top-Down Approaches):(1)Top-Down Approaches,即two-step framework,就是先進行行人檢測,得到邊界框,然後在每一個邊界框中檢測人體關鍵點,連線成一個人形,缺點就是受檢測框的影響太大,漏檢,誤檢,IOU大小等都會對結果有影響,演算法包括RMPE、Mask-RCNN 等。(2)Bottom-Up Approaches,即part-based framework,就是先對整個圖片進行每個人體關鍵點部件的檢測,再將檢測到的部件拼接成一個人形,缺點就是會將不同人的不同部位按一個人進行拼接,代表方法就是openpose、DeepCut 、PAFs。tricks採用多尺度,多解析度的網路結構採用基於Residual Block來構建網路擴大感受野(large kernel, dilation convolution, Spatial Transformer Network、hourglass module)預處理很重要(將人放在輸入圖片的中心,人的尺度儘量歸一化到統一尺度,對圖片進行翻轉、旋轉)後處理同樣重要 三、Single PersonPose estimation2014----Articulated Pose Estimation by a Graphical Model with ImageDependent Pairwise Relations2014----DeepPose_Human Pose Estimation via Deep Neural Networks2014----Joint Training of a Convolutional Network and a Graphical Model forHuman Pose Estimation2014----Learning Human Pose Estimation Features with Convolutional Networks2014----MoDeep_ A Deep Learning Framework Using Motion Features for HumanPose Estimation2015----Efficient Object Localization Using Convolutional Networks2015----Human Pose Estimation with Iterative Error2015----Pose-based CNN Features for Action Recognition2016----Advancing Hand Gesture Recognition with High Resolution ElectricalImpedance Tomography2016----Chained Predictions Using Convolutional Neural Networks2016----CPM----Convolutional Pose Machines2016----CVPR-2016----End-to-End Learning of Deformable Mixture of Parts andDeep Convolutional Neural Networks for Human Pose Estimation2016----Deep Learning of Local RGB-D Patches for 3D Object Detection and 6DPose Estimation2016----PAFs----Realtime Multi-Person 2D Pose Estimation using PartAffinity Fields2016----Stacked hourglass----StackedHourglass Networks for Human Pose Estimation2016----Structured Feature Learning for Pose Estimation2017----Adversarial PoseNet_ A Structure-aware Convolutional Network forHuman pose estimation2017----CVPR2017 oral----Realtime Multi-Person 2D Pose Estimation usingPart Affinity Fields2017----Learning Feature Pyramids for Human Pose Estimation2017----Multi-Context_Attention_for_Human_Pose_Estimation2017----Self Adversarial Training for Human Pose Estimation 四、Multi-PersonPose estimation2016----AssociativeEmbedding_End-to-End Learning for Joint Detection and Grouping2016----DeepCut----Joint Subset Partition and Labeling for Multi PersonPose Estimation2016----DeepCut----Joint Subset Partition and Labeling for Multi PersonPose Estimation_poster2016----DeeperCut----DeeperCut A Deeper, Stronger, and Faster Multi-PersonPose Estimation Model2017----G-RMI----Towards Accurate Multi-person Pose Estimation in the Wild2017----RMPE_ Regional Multi-PersonPose Estimation    這篇是上海交大盧策吾教授專案組的論文,基於Top-Down Approaches。    論文的Motivation就是解決定位誤差和定位框冗餘檢測這兩個問題。引入Google提出的Spatial Transformer Networks,可以使得傳統的卷積帶有了裁剪、平移、縮放、旋轉等特性。    論文中一個實驗:Upper Bound of Our Framework,就是論文直接使用ground truth的人體邊界框,在驗證資料集取得84.2 mAP成績,說明演算法不僅需要提供人體邊界框,第二階段的單人姿態估計效能也需要提高。    腦洞:可以參考MSRA的deformable convolutional network,應該有新的paper。2017----COCO2017 Keypointswinner----Cascaded Pyramid Network for Multi-Person Pose Estimation2017----PyraNet----Learning Feature Pyramids for Human Pose Estimation ---------------------  作者:buling5143  來源:CSDN  原文:https://blog.csdn.net/buling5143/article/details/79466588  版權宣告:本文為博主原創文章,轉載請附上博文連結!