Privately Owned Vehicle Work Group Meeting - 2025/02/10 - Slot 1

[m-zain-khawaja]

Agenda

• Update on Scene3D training
• Update on EgoPath dataset curation
• EgoPath network design discussion
• Update on EgoLanes dataset curation

Discussion

@m-zain-khawaja :

The third phase of Scene3D training using both simulator and real-world data is progressing well. The network loss is converging and the visualizations of the network output are showing promising results in terms of accuracy and refined depth boundaries.

However, although the network is showing good performance for foreground objects as per this example:

The network does not perform well at capturing the depth for certain background objects such as trees, as can be seen in this example:

The likely reason for this is that the SceneContext block was trained as part of the SceneSeg network - which focused on segmentation of foreground objects and did not attempt to classify background objects into different categories (e.g. trees, buildings, poles, traffic lights, traffic signs) - instead combining all of these classes into a single background class. This likely means that the attention of features in the SceneContext block is not focusing on distinguishing features of the background elements of the scene and leading to the less accurate result for Scene3D depth prediction for these types of scene elements. To remedy this, an architectural change has been made wherein a dedicated DepthContext block has been created which should allow the Scene3D network to learn depth features reliably for both foreground and background elements. This new network architecture is now being trained and will likely be completed by the end of this week.

EgoPath Dataset Curation Update

@TranHuuNhatHuy - awaiting update.

@docjag has completed data auditing and has uploaded the completed ROADWorks dataset to Kaggle. @m-zain-khawaja will review and merge his PR with the main branck

Dataset curation tracking

• TuSimple - 6,394 training data samples
• CULane - 15,800 training data samples
• CurveLane - (parsing completed by @TranHuuNhatHuy, data to be uploaded soon)
• BDD100K - 24,500 training samples
• ROADWorks - 4,780 training samples
• Comma2k19 - 31,962 training data samples

Ego Path Network Design

A dedicated group has been setup on Discord to review multiple existing neural network approaches for Lane Detection which can be adapted for EgoPath prediction. A literature review is currently being performed by the work-group members to evaluate various architectural design choices for EgoPath. We expect to have a final network design agreed by the next work group meeting on 17th Feb.

EgoLanes Dataset Curation Update

• TuSimple - 6,394 training data samples
• CULane - 15,800 training data samples
• CurveLane - (to be started)
• BDD100K - (to be started)

@aadarshkt has been assigned the BDD100K dataset parsing task for the EgoLanes Dataset

Attendees

• TBD

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.