CrashSight: A Phase-Aware, Infrastructure-Centric Video Benchmark for Traffic Crash Scene Understanding and Reasoning

CrashSight targets a specific and valuable niche: infrastructure-centric (roadside) traffic safety. While most autonomous driving datasets focus on the 'ego-vehicle' (the car's own sensors), this project addresses the V2I (Vehicle-to-Infrastructure) gap. The defensibility is currently low (score 4) because it is a nascent research project with 0 stars, despite 7 forks indicating early academic interest. Its primary moat is the 'data gravity' of real-world roadside crash footage, which is harder to obtain and curate than standard driving data. However, as a benchmark, its value depends entirely on community adoption; if it doesn't become a standard leaderboard, it will be displaced by similar efforts from larger entities like Baidu, NVIDIA, or academic giants (e.g., Berkeley DeepDrive). Frontier labs are unlikely to build this specifically, but their general-purpose models (GPT-4o, Gemini) are the targets being tested, making the benchmark's longevity dependent on the difficulty of the tasks it presents. The risk of platform domination is medium because hardware providers (like Hikvision or specialized V2X firms) could release much larger proprietary datasets that render this benchmark obsolete.