• [July 20 2021] We have optimized the code and the running speed is much faster. Thank hongfz16 for his useful suggestions.

This repo implements HHA-encoding algorithm in python3. HHA image is an encoding algorithm to make better use of depth images, which was proposed by s-gupta in this paper: Learning Rich Features from RGB-D Images for Object Detection and Segmentation. This algorithm was implemented in MATLAB initially, and I organized a MATLAB version of the code: Depth2HHA-MATLAB. In this repo, I use python3 to get exactly the same result as what MATLAB code does.

All we need is:

1. A depth image
2. A raw-depth image (Option)
3. Camera matrix

If you are confused about depth images or you don't have depth images, please refer to NYU Depth V2 dataset.

I already provide an interface in getHHA.py. The function is named getHHA(C, D, RD). To be detailed, it needs a camera matrix C, a depth image D and a raw depth image RD. RD is used to generate a mask to mark missing information, so it is ok to use getHHA(C, D, D) if you don't have a raw depth image or you just want to do so.

D = cv2.imread(os.path.join(root, '0.png'), cv2.COLOR_BGR2GRAY)/10000
RD = cv2.imread(os.path.join(root, '0_raw.png'), cv2.COLOR_BGR2GRAY)/10000
camera_matrix = getCameraParam('color')
hha = getHHA(camera_matrix, D, RD)

1. I recommend you to use cv2.COLOR_BGR2GRAY instead of cv2.IMREAD_GRAYSCALE to read in the gray image D. If you use cv2.COLOR_BGR2GRAY, you will get exactly the same gray values as MATLAB's imread function does. If you use cv2.IMREAD_GRAYSCALE, you won't.
2. The depth image array passed to function getHHA should in 'meter'. So in my demo code, I divide it with 10000 to modify the unit.
3. Camera matrix C is a 3*3 matrix.

Pictures below are visualizations of:

D:

RD:

HHA from MATLAB:

HHA using D and RD:

HHA using D and D:

If you find this project useful in your research, please consider cite:

@misc{Depth2HHA-python,
    title={{Depth2HHA-python}: Converting depth maps to HHA encodings},
    author={Xiaokang Chen},
    howpublished = {\url{https://github.com/charlesCXK/Depth2HHA-python}},
    year={2018}
}

• Parallel processing