Privately Owned Vehicle Work Group Kick-Off Meeting - 2024/11/11 - Slot 2

[m-zain-khawaja]

Agenda

Provide a summary of the key developments which have already occurred in the PoV work group prior to its formation, including:

• SceneSeg Neural Network
• SceneSeg dataset
• Codebase summary
• Overall development goals and product plan

Attendees

• @m-zain-khawaja (Senior Tech Lead at AWF and Privately Owned Vehicle Work Group Chair)
• Christian John (Chair of Strategic Planning Committee at AWF and President at TierIV)
• David Walmroth (Co-Chair of OpenADKit Work Group at AWF and US Director of Business and Project Development at PIX Moving)
• Tran Huu Nhat Huy (Masters Robotics Student at Tokyo Institute of Technology)

Discussion

• Introductions by attendees
• Run through of agenda items by @m-zain-khawaja
• A presentation was shared by @m-zain-khawaja on the SceneSeg Neural Network, describing the key network features and role of SceneSeg as forming a Safety Shield for autonomous vehicles by detecting all important foreground objects, irrespective of what that object is.
• Tran enquired about the neural network architecture structure, feature backbone design, and how SceneSeg compares with Foundation Segmentation Models such as Segment Anything
• In response, a detailed run-through was presented of the codebase for the SceneSeg Neural Network and its inner workings by @m-zain-khawaja describing the network features, layers, and computational blocks.
• David asked whether the network outputs could be used in dense urban and low-speed driving scenarios, to which @m-zain-khawaja replied that the network's obstacle perception capability should generalise effectively to such scenarios
• Christian John (CJ) enquired about SceneSeg's ability to detect police vehicles and how it compares to the detection capabilities of other companies, to which @m-zain-khawaja replied that police vehicles, and other such rare object types are difficult to classify and form 'long-tail edge case scenarios', and SceneSeg is able to address such scenarios by learning generalisable visual features.
• An overview was presented by @m-zain-khawaja of the overall neural network vision pipeline and high-level ensemble network, called AutoSeg
• @m-zain-khawaja explained that given the training of the network on movable foreground objects, the main detection relates to all movable object categories, and that a separate neural network would be required to complete the vision-pipeline Safety Shield. This second neural network would calculate the true depth of pixels in the scene and use geometric 3D processing to detect the object-ness of every 3D point in the scene.
• @m-zain-khawaja then spoke about the overall development goals and product strategy related to the privately owned vehicle development, alongside the development roadmap and high level plans for Q1 2025.
• Tran said that he would be excited to contribute to this coding development, and enquired as to where he could make a contribution - to which @m-zain-khawaja replied that Tran can begin by conduction a literature review of state-of-the-art lane detection methods covering the key techniques used by researchers, datasets used for training/benchmarking, and methods used to train such networks. Tran said that he would work on this literature review and have it ready for the next WG meeting on Monday next week.

Zoom Meeting Video Recording

Video Meeting Link

Passcode to Access Recording: e8vdY^CL