Instructions on how to run the code

1. Precalculate necessary features from the raw data and store them
   Program: getfeature.cpp
   Input: *.raw, libHandDetect.so
   Output: *.all+
   

     
   					for f in `find dataRecord -name '*.raw'`; do 
   					echo $f; 
   					./getfeature --path $f --handscale 0.9; 
   					done
   					

2. Generate binary tests for the random forests
   Program: genrandfeature.cpp
   Output: testfea.1.dat - not used, testfea.2.dat - uses hand pose, testfea.3.dat - uses raw image data + pose
   
3. Train a forest to classify (vote) clips into action categories
   Program: trainclass.cpp
   Input: testfea.*.dat, *.all+, xml (clips info)
   Output: tree/class/#.tree (# is treeid)
   
4. Train a forest to vote for start/end position of the possible action window containing the current frame
   Program: trainvote.cpp
   Input: testfea.*.dat, *.all+, xml (clips info)
   Output: tree/vote/*-#.tree (* is classid, # is treeid)
   
5. Obtain the votes for each frame (s,e,id) using these two forests
   Program: detectvote.cpp
   Input: tree/class/*, tree/vote/*, testfea.*.dat, *.all+
   Output: *.vote (for each frame, p(class), and for each class a vote)
   
6. Build the voting space for the error-correcting map (in libsvm format)
   Program: trainweight.cpp
   Input: *.vote, xml (clips)
   Output: dclipsvm.dat (rm this file before running), svm/c*/train.param & tset.param (* is classid)
   

     
   					rm dclipsvm.dat
   					./trainweight
   					./trainweight
   					

7. Train the linear error correcting map
   Program: svm/train.sh
   Input: svm/c*/train.param & tset.param (* is classid)
   Output: svm/train.param.model
   

     
   					cd svm
   					for i in {0..numclasses}:
   						do sh train.sh $i;
   					done
   					

8. Convert the learned weights back from libsvm format
   Program: testsvm.cpp
   Input: svm/train.param.model, svm/c*/train.param & tset.param (* is classid)
   Output: svm.w, svm.b
   
9. Perform a test-run of action detection
   Program:testvote.cpp
   Input: svm.w, svm.b, *.vote (for the test data)
   Output: Detection Accuracy/F1 measure
   

1. Q: How to cite the code and the dataset?
   A: Please cite our paper on hand action detection (https://arxiv.org/abs/1606.02031) for the method
   @article { xu2016ECHT,
   author={Chi Xu and Lakshmi Narasimhan Govindarajan and Li Cheng},
   title={Hand Action Detection from Ego-centric Depth Sequences with Error-correcting Hough Transform},
   journal={arXiv preprint arXiv:1606.02031},
   year={2016},
   }
   and for the dataset cite this paper (http://arxiv.org/abs/1701.00903)
   @article { li2017IBGN,
   author={Li Liu and Yongzhong Yang and Lakshmi Narasimhan Govindarajan and Shu Wang and Bin Hu and Li Cheng and David S. Rosenblum},
   title={An Interval-Based Bayesian Generative Model for Human Complex Activity Recognition},
   journal={arXiv preprint arXiv:1701.00903},
   year={2017},
   }

Annotation FAQ:

1. Q: What is the specification file to be used for ANVIL?
   A: Use this file: Updatedcustomized.txt


2. Q: For transition like 12-13-14, 12 and 13 will each share a half of the same pose and 13 and 14 will do the same. How should the pose be split?
   A: The pose should be split as equally as possible so that 12 and 13 will each get a equal half of the shared pose.


3. Q: How to identify the expected starting and ending position of each transition?
   A: The transition should be as compact as possible, which meanly even if there is a long time pause before the transition taking place, the pause should be left out.