During the latest edition of the International Conference on Intelligent Robots and Systems (IROS), at the workshop “Long-Term Perception for Autonomy in Dynamic Human-shared Environments: What Do Robots Need?”, our PI Abhinav Valada, along with Daniele Cattaneo, showcased their work titled “Taxonomy-Aware Continual Semantic Segmentation in Hyperbolic Spaces for Open-World Perception”, for which they were awarded the Outstanding Workshop Presentation Award.
This paper, also authored by our doctoral researcher Julia Hindel, introduces TOPICS, a semantic segmentation model that utilizes class-incremental semantic segmentation to update the model with emerging new classes, but unlike the current state-of-the-art methods, it does not constrain features of the new model to imitate those of the prior model, effectively outperforming previous methods both quantitively and qualitatively.
Their work may present a significant advancement in the development of self-driving vehicles, for which comprehensive scene understanding, capable of adequately learning from the new incoming objects of the open world, is a fundamental need.
We congratulate our researchers on this achievement and look forward to their future work in the field of robotics!