This is the 'half-time'-seminar for WASP industrial PhD student Adam Tonderski.

Abstract:

The reliance on massive amounts of human-annotated data poses a significant challenge for the application of deep learning in the autonomous driving domain. This thesis aims to explore and develop various methods to address this challenge, focusing on three distinct but complementary approaches: (i) enhancing the utility of existing task-specific annotations, (ii) utilizing self-supervision and weak-supervision to train robust foundation models, and (iii) developing techniques to automatically generate annotations.

In the context of these approaches, we present three key works: First, we use standard object detection annotations to train detection transformers to predict objects in future video frames, improving the utility of existing annotations. Second, we propose LidarCLIP, a method that extends the zero-shot capabilities of CLIP to the lidar domain, leveraging self-supervision and weak-supervision. Lastly, as a foundation for all three avenues, we introduce Zenseact Open Dataset (ZOD), a large-scale, diverse multimodal dataset for autonomous driving. Our ongoing research is investigating automatic annotations (approach iii) on ZOD. By integrating these methods, we aspire to reduce the reliance on human supervision in the autonomous driving domain. Further exploration during the seminar will provide valuable insights into the effectiveness and potential impact of our approaches.

The meeting will be in MH:309A, but it will be possible to see the presentation also at 

https://lu-se.zoom.us/j/61259619376?pwd=R2xkNE5NeWZ5VkdLbjJFUW5uL2NQQT09