Skip to content

Latest commit

 

History

History
37 lines (27 loc) · 1.61 KB

dataset_description.md

File metadata and controls

37 lines (27 loc) · 1.61 KB

About the Generated Refresh Dataset

The data is generated using the RefRESH Toolkit. If you want to generate your own data for training, please check the RefRESH repository.

If you want to download the pre-generated data used in the paper at: RefRESH data using BundleFusion.

The calibration and camera pose ground truth are in the info.pkl for each trajectory: e.g. ‘office3/keyframe_1/0000_0100’. The .pkl file was generated by python 2 pickle module. If you load it using python3 please be aware of using python2 pickle module to read :

with open('info.pkl', 'rb') as p:
    # the original pickle is in python2
    files = pickle.load(p, encoding='latin1')

To read the calibration file from the pickle:

calib = files['calib']
# note that the color and depth calibration has very small shift
# (< 0.2 (582.vs 583)). We will only use depth intrinsic since
# it gives very good approximation
Ks = calib['depth_intrinsic']

To read the camera pose from the pickle:

poses = files['pose']

Several Notes:

  • The keyframe indicates the number of frame intervals we used at rendering. In each keyframe sequence, the length of the sequence are 100 at most.

  • The provided optical flow model is backward optical flow, Please be aware of it if you use it to train your dataset. The current toolkit in RefRESH has been updated to support rendering bidirection optical flow. If you need this feature, you can render the data using that tool.