The NVIDIA AI City Challenge: Presenting Our Work at ECCV 2026

At the end of May, I came across the 2026 NVIDIA AI City Challenge and noticed that Track 3 included a task on traffic anomaly reasoning. We had already been exploring this topic, so it seemed like a good opportunity to develop our ideas and compare our approach with others working on the same problem.

I contacted Yan Shi at the University of Washington, whom I had worked with previously, and we decided to take part together. Our collaboration between Bath and Washington eventually led to a runner-up result and the opportunity to present our work at ECCV 2026 in Malmö, Sweden.

Developing our approach

The traffic anomaly reasoning task asked AI systems to understand unusual events in traffic videos: what happened, which road users were involved, and how events unfolded. Relevant details can appear briefly or occupy only a small part of a busy scene, making it important to identify the right evidence before answering a question.

We thought an agentic approach would suit this task. The idea was to build a system that could work out what information a question required and use specialised tools to find it.

This became TAU-Agent. It uses tools that describe video content and track objects to gather relevant evidence, then passes that information and selected video frames to a model that understands both images and language. The model uses these inputs to produce an answer. Our paper explains the approach in detail.

Running the system required paid access to large language model (LLM) APIs. AAPS CDT provided funding for this, which made our participation possible. The IT approval process took around three weeks of the six-week challenge period, leaving us just over two weeks to run our experiments once access was in place.

Testing and results

With access arranged, we focused on testing and improving the system. The results were encouraging, and four days before the deadline we briefly reached first place on the leaderboard. Other teams moved ahead the following day, and we continued refining our approach until the final submission.

We chose to submit our final result close to the deadline, keeping our latest score off the public leaderboard until that stage. We finished in second place. The official results list Team 45 (UOB&UW Team) as runner-up in Traffic Anomaly Reasoning, part of Track 3: Anomalous Events in Transportation, with a mean score of 0.6779.

We were pleased with the result, particularly given the time available for experiments. Taking part gave us a useful way to evaluate ideas from our earlier research and develop them into a working system.

Attending ECCV in Malmö

Following the challenge, I travelled to the European Conference on Computer Vision (ECCV) 2026 in Malmö with Sam Lockyer and presented our work at the AI City Challenge workshop on 8 September. Unfortunately, Yan was unable to attend because of visa issues.

The conference was also an opportunity to learn about research beyond our own project. Sam and I attended talks and spoke with researchers about their work, gaining new insights and different perspectives on problems in computer vision. These discussions gave me ideas to consider in my own research.

I also spoke with people from industry, including Amazon, Google, Meta, ByteDance and Huawei. Learning about their work helped me understand more about how research connects with industry and the opportunities available after a PhD. These conversations were useful as I thought about my future career development.

Reflections

Taking part in the challenge allowed us to build on an existing research interest, strengthen our collaboration and share our work with a wider community. It also gave me experience of managing the practical aspects of a research project, from arranging resources to planning experiments within a limited timeframe.

I am grateful to AAPS CDT for funding the API access, to Yan, Sam and our co-authors for their contributions, and to everyone who supported our participation. Alongside the runner-up result, I particularly valued the ideas and conversations that came from attending the conference.

Our code is publicly available for anyone interested in building on the project.