Privately Owned Vehicle Work Group Meeting - 2024/12/15 - Slot 1

[m-zain-khawaja]

Agenda

• Update on SuperDepth Dataset
• Combining PathDet and DiversionDet networks in AutoSeg
• Data curation update related to new PathDet network

Discussion

Update on SuperDepth dataset download, preparation and curation

• MUAD - 3,912 training data samples
• UrbanSyn - 6,732 training data samples
• KITTI - 4,325 training data samples
• Driving Stereo (found to be not suitable)
• Argoverse - 1,106 training data samples
• MUSES - 2,500 training data samples
• DDAD (unsupported - issue raised with maintainers)
• BOREAS (found to be not suitable)
• GTAV-TeFS (downloaded, parsing in progress)

SuperDepth Training Paradigm

Proposal by @m-zain-khawaja to follow the training methodology of DepthAnything V2 which utilizes a multi-stage training process - the first network stage is trained using 100% simulation data, and this pre-trained network is then used to help generate pseudo-labels on real-world images to help close the sim-to-real domain gap. We can review network performance after the first stage on real-world data to assess the domain gap and use lidar supervision to help enhance pseudo-labels in stage 2.

To further supplement the simulation datasets (MUAD and UrbanSyn) the following GTA V based stereo vision simulation dataset will be parsed and processed, resulting in an additional 20K simulation-based training samples.

PathDet dataset curation

• TuSimple - 6,394 training data samples
• CULane - (downloaded, parsing in progress)
• CurveLane
• BDD100K
• ROADWorks

Pull request has been successfully merged for the parsing of the TuSimple dataset as part of PathDet dataset curation by @TranHuuNhatHuy

Work is now underway for processing and parsing CULane, BDD100K and ROADWorks - with aim for @TranHuuNhatHuy to complete CULane parsing this week.

Attendees

@m-zain-khawaja (Senior Tech Lead at AWF and Privately Owned Vehicle Work Group Chair)
@mitsudome-r (Chair of Autoware Technical Steering Committee)
@TranHuuNhatHuy (Masters Robotics Student at Tokyo Institute of Technology)
Narendar Selvaraj (Senior Member of Technical Staff at AMD)
Lucus Xingang Liu (Autocore.ai)
Siddharth Das (ADAS Engineer at Canoo)
Devang Marvania (Lead Engineer - User Experience, Calibration at Gentherm)

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.