KEYWORD |
DAUIN - GR-23 - VANDAL - Visual and Multimodal Applied Learning Lab
Towards enhanced anomaly segmentation in driving scenarios: developing a multi-modal synthetic dataset with CARLA
keywords SYNTHETIC 3D MULTIMODAL DATASET, ANOMALY DETECTION
Reference persons ANDREA BOTTINO, CARLO MASONE, TATIANA TOMMASI
External reference persons Leonardo Vezzani
Research Groups DAUIN - GR-02 - COMPUTER GRAPHIC AND VISION GROUP - CGVG, DAUIN - GR-23 - VANDAL - Visual and Multimodal Applied Learning Lab
Thesis type RESEARCH
Description
The ability to detect anomalies in driving scenes is crucial for the safe deployment of autonomous vehicles. Current benchmarks for anomaly segmentation, such as FishyScapes, SegmentMeIfYouCan (SMIYC), and RoadAnomaly, offer some insights into this challenge. These datasets primarily focus on anomalies as new, unseen objects placed on the road, represented through monocular camera images. However, this narrow definition fails to capture the broader and more complex nature of anomalies. Anomalies can also include familiar objects in abnormal configurations (e.g., a fallen tree or a car in an unusual position), as well as environmental or contextual irregularities. This thesis proposes to address these limitations by creating a new synthetic dataset using the CARLA driving simulator. In particular, we aim to:
Support multi-modal anomaly detection methods, by collecting a dataset with multiple simulated sensors (RGB cameras, depth cameras, Lidar, radar, event cameras)
Support the investigation of these methods across different illumination and weather conditions. This raises some challenges for the realism of the simulated sensors (e.g., implementing the effect of rain on the lidar sensor)
Integrate both static and moving anomalies from a diverse set of categories (animals, objects)
Include anomalies that are not just new object categories. For example, an anomaly could also be a common object category but in a strange configuration (e.g., a fallen tree or a car in an accident).
The development of a well defined, customisable and possibly automatic pipeline to create short simulation episodes in CARLA, to streamline the data collection.
Required skills Computer graphics, programming in Python, and optional knowledge of Unreal Game Engine
Deadline 29/01/2026
PROPONI LA TUA CANDIDATURA