implement some algorithms of 6d pose estimation

• linemod
• linemod levelup
• linemod ros
• latent class hough tree(LCHF)
• 3d convex segmentation

note: LCHF is too hard to tune params so I'm not going to implement it...
a better choice: patch_linemod

files: params/ pysixd/ t_less_toolkit/ tools/
copied from sixd_toolkit
deal with model reading/rendering, datasets reading and evaluation

get dataset under 6DPose folder using following cmd

wget -r -np -nH --cut-dirs=1 -R index.html http://ptak.felk.cvut.cz/6DB/public/

install opencv3 with contrib rgbd module
install pybind11

pip3 install -r requirements.txt

in target folder:
mkdir build
cd build/
cmake ..
make

in 6dpose folder, if use pybind:
pip3 install target_folder/

python3 target.py

Codes in linemod.py will train and detect objects in downloaded dataset.
Refer to opencv linemod and ork linemod src

Linemod suffers a lot from scale and occlusion problems, to deal wtih them:

1. set low response closer to zero;(idea got from focal loss)
2. use depth histgram and 1D nms to find some primary scales

ori linemod can't run while our features are more than 64, due to 8bit SSE. after changing to 16bit SSE and preventing overflow, we can have up to 8192 features now.

Chinese blog about the ideas

some tests

run linemod with ros for real test, not done yet

• linemod patch version
• hough forest
• meanshift for leaf mode
• icp for pose refine
• train and test

Chinese blog 0
Chinese blog 1

segment 3d cloud to convex part, core func have been done, try to deploy now

NOTE: To use pybind, static super4pcs must be built with flag -fPIC. Build super4pcs using following lines:

mkdir build
cd build
cmake .. -DCMAKE_CXX_FLAGS="-fpic"
make
sudo make install

separated repo for deployment

Chinese blog