Activity Page


On this page I intend to keep an updated track of what I am working on at the moment and the next steps in the foreseeable future.

State-of-the-art:
  • Entry 101 - Concluded a formal state-of-the-art report (even though I will keep on studying for new methods and approaches as the the project evolves).

Descriptor implementation and testing:
  • Entry 201 - After several "back and forth" due to technical problems, I have finally managed to get the SURF (Speeded Up Robust Feature) descriptor working (detection, extraction and comparison methods).
  • Entry 202 - Currently working on the HOG (Histogram of oriented gradients) descriptors. Almost done, just need to re-think on how best to use the appropriate metric, regarding histogram comparison. Also, I should think about on how to integrate this module onto the existing code, since with HOG there's no such thing as a "keypoint" or "interest point".
  • Entry 203 - The use of HOG features ended up working great with a good execution time and achieving high rates of recognition.
  • Entry 204 - Succefully imlemented the use of Grid Points using the SURF descriptor ("forcing" the points' coordinates rather than having a detection stage).
  • Entry 205 - A lot of memory leak problems were detected on the SURF descriptor (maybe that's why its executable is so slow). All the problems ended up being corrected.
  • Entry 206 - Making a few changes on all of the descriptors to deal with oclusion situations. Until the oclusion ratio reaches a defined threshold, the object model is updated. From then on, no changes are made. Also, the detected keypoints are only  considered if they are located on the Ellipse Model of the object. Needs further work.
  • Entry 207 - A lot of memory leaking problems corrected.
  • Entry 208 - Using bag-of-visterms approach. A dictionary was created using random images and image sequences, and then all that was needed to do was a to run a query and evaluate the histogram.
  • Entry 209 - Started using the FIRST descriptors. A very limited API makes it hard to integrate it with the tracking algorithm. Execution time results are very good, but tracking accuracy was disappointing.
  • Entry 210 - Started integration of all the work done towards descriptors and appearance models to integrate them on a common evaluation Framework. The results will be posted on the master thesis.