Yuk Hwa-chui, a student in the integrated master’s-doctoral program at Kookmin University’s Graduate School of Automotive Mobility (advisor: Professor Lee Sanghun), delivered an oral presentation of her paper at the ECCV 2026 Safe and Defensive Autonomous Driving (SDAD) Workshop held in Malmö, Sweden, on September 9, and received an Honorable Mention for Best Paper.

The award-winning paper, titled “Seeing the Evidence, Not Just the Answer: Training-Free SalMask Evidence Routing for Traffic Accident Video Understanding,” is a study that explores how vision-language models can utilize visual evidence to understand traffic accident videos. The Best Paper Honorable Mention is a special recognition awarded to outstanding papers nominated for the Best Paper Award; at this workshop, it was awarded to two papers nominated for the Best Paper Award.
 
ECCV (European Conference on Computer Vision) is a leading international conference in the fields of computer vision and machine learning; ECCV 2026 was held in Malmö, Sweden, from September 8 to 12. The SDAD workshop is an academic conference focused on safe and defensive autonomous driving, discussing topics such as the recognition of potential hazards in driving scenarios, responses to uncertain situations, and methods for evaluating defensive driving. Eighteen papers were ultimately accepted for this year’s workshop, four of which were selected for oral presentation.
 
In this research, Yook Hwa-chui, a Ph.D. candidate in the integrated master’s-Ph.D. program, proposed the SalMask Evidence Routing (SalMask-ER) method to enable Vision-Language Models (VLMs) to utilize scenes and objects—which serve as the basis for decision-making—more accurately, rather than relying solely on linguistic reasoning. SalMask-ER is a training-free method that does not require additional model training; it simultaneously feeds the model local visual information reflecting the overall video context, accident-related objects, and the driver’s area of attention.
  
Experimental results showed that the proposed method improved the accident video question-answering accuracy of the Qwen2.5-VL-3B model from 51.12% to 67.33% on the VRU-Accident benchmark, while the Qwen3.5-4B model achieved an accuracy of 71.30%. This award is recognized as a research achievement that demonstrates the importance—not only of answer accuracy in understanding traffic accident videos—but also of how the model judges and explains the accident situation based on specific visual evidence.