image: (half of the obj is occluded)

ori detector: threshold 84 + nms

set low response closer to 0

scale experiment

Firstly we train templates by rendering img from 600mm, while obj is about 1000mm in scene. As is expected, our ori detector fails:

ori low to 0 version(simi drops off course)

scale template at each match position, good but 10 times slower

histogram + 1D nms to find primary depth

As we can see, the template is trained from 600mm. We use histogram + 1D nms to find possible depth in scene, in this case we find 5 possible depths, and successfully, 1000mm is one of them. Matching time is about 60ms now.

todo: more than 64 features. need to modify similarity(local), and addSimilarities(delete 8u 8u), and distinguish them because use 16 sse may be slower than 8 sse
DONE

Detector: n clusters params
assign area to feature and hist area info(how many feats in one area) to templ in cropTemplates
no need for hist, hist when matching
area pyrDown: clsuter after pyrDown
similarity_64 & similarity: add to different dst(dst become a vec now) according to featrue area
active specific area by the thresh of simi.
class Match: pass active areas to it.
func match: add a min num of active part param

!!!!!! num_feat and clusters /4 when pyrDown, may cause problems?
well, back to /2, /4 or /2, it's a question.