Vectornet lstm evaluator is is based on a GNN encoding network taking environments and traffic agents as nodes. Comparing with semantic map CNN model, it has less parameters to learn and better performance.
Please refer the online inference vectornet lstm evaluator.
[1]: Gao, Jiyang, et al. "Vectornet: Encoding hd maps and agent dynamics from vectorized representation." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.