readme.md

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

todo:
spread step same as stride or 2:1?
PSO or some other method rather than ICP

may implement PSO:
Correspondence-free pose estimation for 3D objects from noisy depth data

1D nms to select primary depth, or cut to some primary depth boxes?

nms for accurate edge

with selective search? selective search in depth modality?

2bit label _mm_max_epi8