TO DO.txt

PREPARATIONS:
=============
* Download video databases
* Read frames from video data
* Background calculation and mask generation for each video
* Download library of background separation
* Tag videos for features, find best combination <hoc,dist, mu> for that and use a supervised learning to predict best combination


HOC:
====
* Color Histogram Peak Selection
* Using only bins from the full set that foreground is not zero in them
* HSV Space (8x8x8), (8x8x4)
* Add V channel of HSV space to RGB to have RGB-V space
* JND HOC
* Gridding using adaptive vertical projection
* Gridding only vertically (2x1, 3x1, 5x1)
* Use color moments
* use the rank of HOC bins instead of their values (value = descending sort index) and normalize them by deviding them by n(n+1)/2
* Implement corellograms


DISTANCES:
==========
* Statistical Distances of two histogram (mean and variance based) -> weigthed mean variance, bhattacharyya distance, 
* Pixel-wise Template Matching


MODEL UPDATE:
=============
* Model update: Grid based update using beta distribution


VISUALIZATION:
==============
* Visualize sliding window with box color indicating similarity to template and showing the ground truth position
* Similarity map and ranking works as alpha blend of the color
* Visualize the similarity of BB boxes and template as heat map, the similarity is the color of BB center


FURTHER PROCESSING:
===================
* Compare Each Distance measure over different HOC schemes, for intra and template
* Calculate inter to intra ratio for each combination (the smaller, the better)
* Different HOC with Best Param, Similar Dist
* Same HOC with Different Params, Similar Dist
* Same HOC with Same Params, Different Dist
* Same HOC with Same Params, Different Dist, Different Gridding
* Plot: (Score vs. Number of Bins) --> different chart for each HOC, each line is (dist,mu)
* Plot: (Score vs. Grid Size) --> each line is (hoc,dist,mu,param)

JOURNAL 2 ALGORITHM:
====================
calculate all boxes possible for this video
for all frames
... for all bounding boxes
...... for all HOC+grigging types
......... for all model update models
............ for all distances
............... calc dist for (hoc+gridding, dist, mu)
............... sort all box distances
............... keep the best box
............... keep the rank of ground truth box
visualize best boxes
calculate average ranking of best boxes as a measure of good ness of (hoc,dist,mu)