James Steven Supančič, III, Ph.D.

• J. Supancic, D. Ramanan. " Tracking as Online Decision-Making: Learning a Policy from Streaming Videos with Reinforcement Learning" International Conference on Computer Vision (ICCV) Venice, Italy, 2017.

• B. Kong, J. Supancic, D. Ramanan, C. Fowlkes "Cross-Domain Forensic Shoeprint Matching" British Machine Vision Conference London, United Kingdom. 2017.

• I recently completed my Ph.D. in Computer Science at the University of California, Irvine.
• You can find my dissertation here Long-Term Tracking by Decision Making

1. Our survey and evaluation of hand datasets and pose estimation methods is on arXiv
2. J. Supancic, G. Rogez, Y. Yang, J. Shotton, D. Ramanan. "Depth-based hand pose estimation: methods, data, and challenges " arXiv preprint arXiv:1504.06378 2015.
3. • Deep Convolutional Network w/ Prior Bottleneck source code and a pre-trained network
   • Other methods, specifically including Dual-Tree accelerated 1-NN
   • Modified LibHand w/ Inverse-Kinematic Annotation
   • Dataset Link
4. Links to datasets mentioned in the paper: KTH , LISA , ASTAR , MSR , NYU , ICL , FORTH , Dexter , ETH-Z , Max-Planck-Gesellschaft ( Synthetic , Real ), HandNet , FingerPaint

I'm organizing a workshop and challange associated with CVPR-2015. More information can be hound here .

Hand pose estimation from an egocentric view using random cascades with synthetic depth data

• G. Rogez, M. Khademi, J. Supancic, J. Montiel, D. Ramanan. "3D Hand Pose Detection in Egocentric RGB-D Images" Workshop on Consumer Depth Cameras for Computer Vision, European Conference on Computer Vision (ECCV), Zurich, Switzerland, Sept. 2014. PDF

Our paper, "Self Paced Learning for Long-Term Tracking" was accepted for publication at CVPR 2013. The paper presents a novel technique for adapting an appearance model for long term tracking.

• Project Source Code (Release Version, WIP)
  1. See it's README.txt for instructions.
  2. Feel free to email me if you have problems.
• Self-paced learning for long-term tracking (Paper in PDF form)
• Videos and Ground Truth (Data Set)
• Output Tracks (CSV: x1, y1, x2, y2, cost). If x1 == NaN then the tracker said the target was occluded or out of frame.
  1. Offline Output Tracks (CSV format)
  2. Online Output Tracks (CSV format)

• J. Supancic, D. Ramanan. "Self-Paced Learning for Long-Term Tracking" Computer Vision and Pattern Recognition (CVPR), Portland, OR, June 2013.

The imageeq function, which follows, equalizes a floating point image before displaying it. It is much better than MATLAB's imagesc for visualizing depth data.

void imageeq(const char* winName, cv::Mat_< float > im)
{
    // compute the order statistics
    vector< float > values;
    for(int rIter = 0; rIter < im.rows; rIter++)
        for(int cIter = 0; cIter < im.cols; cIter++)
                values.push_back(im.at< float >(rIter,cIter));
    std::sort(values.begin(),values.end());
    auto newEnd = std::unique(values.begin(),values.end());
    values.erase(newEnd,values.end());
    
    // compute an equalized image
    Mat showMe(im.rows,im.cols,DataType< uchar >::type);
    float oldQuant = 0;
    for(int qIter = 1; qIter <= 256; qIter++)
    {
        float quantile = static_cast< float >(qIter)/256;
      
        float thresh_low = values[oldQuant*(values.size() - 1)];
        float thresh_high = values[quantile*(values.size() - 1)];
        //printf("q = %f low = %f high = %f\n",quantile,thresh_low,thresh_high);
        for(int rIter = 0; rIter < im.rows; rIter++)
            for(int cIter = 0; cIter < im.cols; cIter++)
            {
                float curValue = im.at< float >(rIter,cIter);
                if(curValue <= thresh_high && curValue >= thresh_low)
                    showMe.at< uchar >(rIter,cIter) = qIter-1;
            }
        
        oldQuant = quantile;
    }
    
    imshow(winName,showMe);
}

The above figure contains the inverse HOG visualization of the HOG features in a picture of me. I am a first year PhD student at UC Irvine. I work in computer vision. I'm interested in how we might exploit temporal data (tracking) and depth data (e.g. Kinect) in semi-supervised learning to create effective detectors.

1. backwm - A program that allows application to use backend displays directly when you are using Xdmx. Last Updated (10.6.2006).
2. a chromium installation. Last Updated (10.6.2006)
3. a small utility that helps create interpose librarys (28.6.2006)
4. a tool that can be used to generate library call logs (28.6.2006)
5. a interpose library used to log all calls made to Xlib (28.6.2006)
6. a chromium SPU that allows the direct use of bakcend hardware in a DMX environment (13.10.2008)