jindongwang / activityrecognition

Resources about activity recognition-行为识别资料

Hi 👋, I'm Jindong Wang

• 😄 Hi there, this is Jindong Wang. I am a Senior Researcher at Microsoft Research Asia (MSRA).
• 🔭 My research interest includes robust machine learning, transfer learning, out-of-distribution generalization, machine learning, and other applications.
• 👯 I am open to collaboration, feel free to contact me via Email (👈)!
• ⚡ Please check my homepage for my CV and latest update!

activityrecognition's Issues

FFT特征提取问题

你好，刚接触信号处理，看了你的代码，对于频域特征提取几点问题想要请教：
1、你有说过窗口大小必须是2的指数倍，傅里叶变换才能才能顺利进行，可在使用python的np.fft函数时并没有体现这一点
2、你在计算形状统计特征的均值时，freq_spectrum分量乘的系数是0,1,2,...n-1,是不是应该乘以1,2,3,,,,n

tsfeatures時間序列特徵提取問題

大神您好，
目前在R語言上有search到這個套件，主要也是做時間序列上的特徵提取，不知道您這個套件與他是否相似？
https://github.com/robjhyndman/tsfeatures

发现个HMDB数据库

链接在http://serre-lab.clps.brown.edu/resource/hmdb-a-large-human-motion-database/#dataset

`fft_shape_kurt`函数没有加return

def fft_shape_kurt(self):
    shape_mean = self.fft_shape_mean()
    np.sum([np.power((x - shape_mean), 4) * self.freq_spectrum[x] - 3
                for x in range(len(self.freq_spectrum))]) / self._freq_sum_

test函数有误？

feature_core.py文件中的test()函数有误，将数组a安排成10行1列的数组后，计算数组a的fft只会直接返回a，因为此时是给数组a中每个元素计算一次fft，而不是给全部元素计算一次fft

请问`fft_shape_kurt`为什么要减3呢？

def fft_shape_kurt(self):
    shape_mean = self.fft_shape_mean()
    return np.sum([np.power((x - shape_mean), 4) * self.freq_spectrum[x] - 3
                for x in range(len(self.freq_spectrum))]) / self._freq_sum_

Some questions regarding the source code

To the author of the paper "Stratified Transfer Learning for Cross-domain Activity Recognition.":

I come across your work in cross domain adaptation for activity recognition and find it very interesting to read. I was just wondering, if you have the source code also available in python or pytorch?

Other questions regarding the paper:
1.) Does the source and target domain have the number of samples for each classes?
2.) Should the source and target domain have the same amount of samples?
3.) Do you update the mmd_loss for each batch or the entire dataset?
4.) Can you maybe also share the data with me? I can't find the matlab matrix for dsads.mat.

Thanks for your reply!

Best regards,
Biying

why not see the data_combined？

Hi,where are the data_combined of the line of FeatureExtraction2?

Reproducing cross-dataset STL

Hi, how can I reproduce the result for cross-dataset as shown in your STL (percom 18) paper? Is the test run for cross-dataset already in the code somewhere? I coded myself using STL function (in STL.m file), trying DSADS --> PAMAP, the accuracies were always <30% (while it is shown 37.83% in the paper). Here is my attempt:

cross_dsads = load('C:\dataset\crossposition-activity-recognition\cross_dsads.mat');
cross_dsads = cross_dsads.data_dsads;

cross_pamap = load('C:\dataset\crossposition-activity-recognition\cross_pamap.mat');
cross_pamap = cross_pamap.data_pamap;

Xcross_dsads = cross_dsads(:,1:81); %features for torso
ycross_dsads = cross_dsads(:,end);
Xcross_pamap = cross_pamap(:,82:162); %features for chest
ycross_pamap = cross_pamap(:,end);

acc_stl = STL(Xcross_dsads, ycross_dsads, Xcross_pamap, ycross_pamap,30);
fprintf('Acc: %.2f\n',acc_stl);

Do you think I missed something?

In addition, feature_norm is not defined in the "demo.m", when I tried using opportunity dataset.

Thanks,
Aria

数据格式问题

王博士，你好

   6轴陀螺仪和加速度计提取特征，数据格式应该是怎么样的呢？

ActivityRecognition and ActionRecognition

Hello, and I'm sorry for disturbing you.But I really want to know the difference between Activity Recognition and Action Recognition? Could you please show me some lights on?
Thank you so much!~~~