Montag, 5. Juli 2010

Cascaded Confidence Filtering for Improved Tracking-by-Detection




S. Stalder, H. Grabner, and L. Van Gool

In Proceedings ECCV 2010 [paper][slides][bibtex][scovis dataset sample (4000 images)]

We propose a novel approach to increase the robustness of object detection algorithms in surveillance scenarios. The cascaded confidence filter successively incorporates constraints on the size of the objects, on the preponderance of the background and on the smoothness of trajectories. In fact, the continuous detection confidence scores are analyzed locally to adapt the generic detector to the specific scene. The approach does not learn specific object models, reason about complete trajectories or scene structure, nor use multiple cameras. Therefore, it can serve as preprocessing step to robustify many tracking-by-detection algorithms. Our real-world experiments show significant improvements, especially in the case of partial occlusions, changing backgrounds, and similar distractors.

4 Kommentare:

  1. I've read your papper and found it very interesting. I've one question about your object detection confidence map. How do you compute it ? Is every pixel of this map corresponds to the raw output of the hog svm detector ? And how do you integrate the scale in this map ?

    Regards,

    Jaonary

    AntwortenLöschen
  2. Thanks for your interest in our work.
    You are right, every pixel of this map corresponds to the raw output of the HoG SVM detector. The scale of the persons is assumed to be known in advance. More precisely, we estimate the ground plane on which persons walk. With the assumption of a flat ground plane, it is easily possible to infer the height of persons at each location in the image.

    If you have further questions, please do not hesitate to contact me again.

    AntwortenLöschen
  3. Hi, Severin, your tracking by detection work is very impressive, and I think the performance of the detection parts could be improved by GPU acceleration to achieve real-time performance with some ROI restriction, actually the CUDA based HOG detection is implemented in opencv 2.2, you can try it.
    In addition, would you like to share a copy of your presentation and code with me if it's convenient. My email is: qqsongzi@gmail.com
    BRs,
    Qingsong

    AntwortenLöschen
  4. Hi Qingsong,
    Thanks for your interest in our work.
    The presentation can be found here:
    http://www.vision.ee.ethz.ch/%7Esstalder/presentations/ccfSlideShow.pdf

    We are currently working on publishing the source code of our implementation. It should be ready by mid-April. I will let you know. Best regards, Severin.

    AntwortenLöschen