meiqua / 6DPose

implement some algorithms of 6d pose estimation

6DPose

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

prerequisite

pysixd

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

dataset

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/

library

install opencv3 with contrib rgbd module
install pybind11

pip3 install -r requirements.txt

steps

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

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

python3 target.py

linemod

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

result

linemodLevelup

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

linemod ros

run linemod with ros for real test, not done yet

latent class hough forest --- LCHF

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

Chinese blog 0
Chinese blog 1

3d convex segmentation

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

result of cloud_seg.py

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