Privately Owned Vehicle Work Group Meeting - 2024/12/02 - Slot 1

[m-zain-khawaja]

Agenda

SuperDepth Dataset Update

• Update on dataset curation for SuperDepth network
  • new LIDAR depth completion algorithm implementation in a separate helper class, based on In Defense of Classical Image Processing: Fast Depth Completion on the CPU for creation of dense ground truth depth map from sparse LIDAR depth data projected on a virtual camera
  • implementation of sparse stereo matching on CPU using OpenCV Block Matching algorithm to create sparse supervision for datasets where stereo images are available

Update on dataset download, preparation and curation

• MUAD - 3912 training data samples
• UrbanSyn - 6,732 training data samples
• KITTI - 4,325 training data samples
• Driving Stereo (in progress)
• Argoverse (to be started, download complete)
• MUSES (to be started, download complete)
• DDAD (currently downloading)
• BOREAS (currently downloading)

Attendees

• @m-zain-khawaja (Senior Tech Lead at AWF and Privately Owned Vehicle Work Group Chair)
• @mitsudome-r (Chair of Autoware Technical Steering Committee)
• Tran Huu Nhat Huy (Masters Robotics Student at Tokyo Institute of Technology)
• Rineez (Chief Technology Officer at Livares Technologies)
• Siddharth Das (ADAS Engineer at Canoo)
• Devang Marvania (Lead Engineer - User Experience, Calibration at Gentherm)

Discussion

• @m-zain-khawaja provided an update on the dataset curation related to the SuperDepth network
• @m-zain-khawaja explained the new helper classes which were created to aid in dataset curation tasks such as lidar_depth_fill.py, stereo_sparse_supervision.py, and height_map.py
• @m-zain-khawaja explained the purpose of each helper class and in particular, explained the algorithm design behind LIDAR depth fill algorithm as well as the stereo sparse matching algorithm being implemented in the helper classes.
• @TranHuuNhatHuy then presented the latest updates on the Lane Detection literature review, including the datasets that could be used for training, these were TuSimple, CULanes, CurvLanes, BDD100K, and LLAMAs
• @m-zain-khawaja commented on the importance of having a single label format for ground truth data, and that we should think about methods to unify the ground truth across the datasets
• @m-zain-khawaja also commented on LLAMAS dataset and showed, with reference to a presentation by Bosch (who created this dataset), that the dataset has limitations on account of the automatic labelling methodology of projecting lanes labelled in an HD map onto an image plane
• It was agreed that we would not use the LLAMAS dataset for training and would rely on the other datasets instead.

Zoom Meeting Video Recording

Video Meeting Link
Please contact the work group lead (@m-zain-khawaja ) to request access to a recording of this meeting.