the newly implemented baseline + DCT does not improve miss rate as expected

[LevinJ]

The detailed issue description is as below:

1. Baseline detector performace on Inria dataset is 20%
2. Have the input data decorrelated by adding DCT channels is supposed to reduce miss rate by about 3%, which is about 17%.
3. But the actual resut obtained from baseline+DCT is 42%.

[LevinJ]

There are quite a few areas which I could look into and figure out why the new detector has this performance gap with what is obervred in “Benenson etc., 2014 EECV” paper. http://rodrigob.github.io/documents/2014_eccvw_ten_years_of_pedestrian_detection_with_supplementary_material.pdf

1. The timing of performing DCT transformation
   Currently DCT transformation is performed right after the 10 channels (LUV + 6 HOG s + 1 Gradient) are computed and before the shrinking of these 10 channels. maybe the DCT channels needs to be done after the shrinking.
2. Variants of DCT transformation
3. Adboost training process
   Maybe the trainig process needs to be accordingly adusted after adding the DCT channels.

[LevinJ]

Have made several attempts to fix this issue:

1. Adjust the timing of performing DCT by doing it after the shrinking, but this turns to worsen the situation. see here for details. Need to try performing DCT after feature shrinking, as opposed to before the feature shrinking #4
2. Perform DCT on individual channel, insead of the whole combined big channel. this does improve the rate by about 2%, see here for details. Need to try performing DCT separately on each channel image #5
3. Debuged into the Adboost training process, and didn't find anything that might cause DCT method not to work. Need to add more traces for adaboost training process to help figure out why dct method does not work as expected #2

Close the exploration on this issue for now, till I have better ideas where to look for the cause.