Privately Owned Vehicle Work Group Meeting - 2024/11/25 - Slot 1

[m-zain-khawaja]

Agenda

SuperDepth

• Review of dataset preparation for SuperDepth neural network including datasets which have been selected for curation, methods used for dataset preparation and progress update on dataset creation.
• Update on literature review for lane detection related to LaneDet development.

Attendees

• @m-zain-khawaja (Senior Tech Lead at AWF and Privately Owned Vehicle Work Group Chair)
• @mitsudome-r (Chair of Autoware Technical Steering Committee)
• Simon Thompson (Director of Engineering, Research at Tier IV)
• Hiroshi Igata (Lead of ODD Work Group at AWF)
• Alexander Korvyakov (Senior Product Manager - Autonomous Vehicle Foundation at MOIA, part of VW Group)
• Tran Huu Nhat Huy (Masters Robotics Student at Tokyo Institute of Technology)
• Krishna Thallapalli (Technical Specialist at Hexagon Manufacturing Intelligence, India)

Discussion

• @m-zain-khawaja presented the latest updates on developments of the SuperDepth network, specifically related to data curation for training
• the training datasets being used for SuperDepth include simulated data, lidar + monocular camera data, and lidar + stereo vision data
• @m-zain-khawaja showed the input data and ground truth data requirements for the SuperDepth network, including images, depth boundaries, metric heigh-map and metric depth-map
• a code walkthrough was performed of the pipeline to process the raw data from MUAD/UrbanSyn simulator datasets to create the required inputs for the SuperDepth network with a focus on the key functions to create the ground truth depth map, create a depth boundary mask, simulate spare stereo supervision, and create a ground truth height map
• @mitsudome-r enquired about how an obstacle perception pipeline could leverage vision-derived depth data for robust scene perception, highlighting the importance of being able to remove noise whilst not filtering out true-positive obstacle detections
• @m-zain-khawaja replied that existing methods for pointcloud processing are able to reliable tackle this and @m-zain-khawaja expressed confidence that if the derived depth output of SuperDepth is on-par or better than existing computer vision approaches such as Semi-Global Batch Matching (SGBM), then it should be able to be processed for reliable obstacle perception, citing the example of (Ambarella group's stereo vision obstacle perception system)[https://www.ambarella.com/blog/how-a-car-sees-in-3d/]
• Alex from MOIA, part of Volkswagen Group introduced himself to the work group and spoke about MOIA's work to develop autonomous ride share platform in collaboration with Mobileye and MOIA's focus on L4 autonomy
• Tran shared that he aims to make progress on the LaneDet literature review and intends to join Slot 2 to share this update.

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.