Course talk:CPSC522/RoadgraphGraphNeuralNetworks

Overall, the project utilizes GNN for Roadgraph encoding which is very interesting and very relevant to the course. Using adjacency of the road graph is quite an interesting approach. Something more visual could potentially offer insight into why GCNs work better. I would like to see more analysis/insight as to why GCNs worked well. Minor Errors:

• graph based -> graph-based self driving -> self-driving (many more like this)
• behavior-> behaviour
• for each lane segment according [to] the formula

Still, this is very impressive for an individual project, good work!

The motivation was well-explained and it was clear how each of the architectures differed from one another! The figure at the end also nicely visualized all of the results!

As mentioned in the below critique, some additional visuals could help make things clearer such as including the ones from the presentation.

It would be nice to have some more elaboration on each of the architectures like the advantages of graph convolutional and graph attention networks and which ones you thought would perform better.

It could also be mentioned explicitly that better performing architectures will have higher mAP.

Critique

1) Overall, this project report is well-written and relevant for the course. The writing is clear and the English is good, making it easy to understand for the intended audience. The page is written at an appropriate level for the course and the formalism used, including definitions and mathematics, was well chosen. Overall, the content is informative and well-structured, providing a good high-level overview of roadgraph graph neural networks and how this report addresses issues with existing solutions.

2) It would be great if you could add a corresponding visual to the dataset description. Maybe something from https://www.argoverse.org/?

3) Possibly a little nitpicky but why does the lane segment have to specifically be within 2m of the final ego vehicle position? Is this a common cutoff in literature?

4) Although the formalisms are generally easy to interpret given the context, I think it would be even better (considering the fact that readers of this course could potentially be from diverse backgrounds) if you could just introduce some of the variables before introducing the full function (for GCN and GAN).

Minor grammatical errors

1) Self driving cars have the potential to revolutionize transportation by reducing time spent driving and traffic *fatalities.

2) One key aspect of this technology is motion forecasting (no comma here) which, along with perception and motion planning systems*, *forms the pillars of autonomous vehicle software.

3) In addition, the motion of agents is influenced by the motions of other agents as they aim to safely navigate the driving environment*.