During the development of ADAS systems, one of the key problems is the effective identification of rare, so-called long tail or black swan events, and that such events are represented in as large a number as possible. Such events are, for example, road accidents, abnormal vehicles, etc. appearance on the roads, and the so-called near-miss situations. In the case of large data sets, finding such events is time-consuming, the task is to automatically identify such events (the list above is not complete) using, for example, CLIP or similar vision-language transformers. Publicly available datasets as examples (the corresponding devkits): https://github.com/nutonomy/nuscenes-devkit and https://github.com/yizhou-wang/cruw-devkit.

During the development of ADAS systems, one of the key problems is the effective identification of rare, so-called long tail or black swan events, and that such events are represented in as large a number as possible. Such events are, for example, road accidents, abnormal vehicles, etc. appearance on the roads, and the so-called near-miss situations. In the case of large data sets, finding such events is time-consuming, the task is to automatically identify such events (the list above is not complete) using, for example, CLIP or similar vision-language transformers. Publicly available datasets as examples (the corresponding devkits): https://github.com/nutonomy/nuscenes-devkit and https://github.com/yizhou-wang/cruw-devkit.