General

1. We would like to participate. What do we need to do?

This time the data access information is shared on each track’s description page under CHALLENGE tab. Please find the instructions in there. 

2. I am interested only in submitting a paper but not in the Challenge. Can I do that?

Yes. Please make sure to submit your paper by the submission deadline.

3. How large can a team be?

There are no restrictions on team size.

4. Are we allowed to use other external data/pre-trained models?

External dataset or pre-trained models are allowed only if they are public. Teams that wish to be listed in the public leader board and win the challenge awards are NOT allowed to use any private data or private pre-trained model for either training or validation. The winning teams and runners-up are required to submit their training and testing codes for verification after the challenge submission deadline in order to ensure that no private data or private pre-trained model was used for training and the tasks were performed by algorithms and not humans.

5. What are the prizes?

This information is shared in the Awards section.

6. Will we need to submit our code?

Teams need to make there code publicly accessible to be considered for winning (including complete/reproducible pipeline for mode training/creation). This is to ensure that no private data is used for training and the tasks were performed by algorithms and not humans and contribute to the community.

7. How will the submissions be evaluated?

The submission formats for each track are detailed on each track’s description page under CHALLENGE tab.

The validation sets are allowed to be used in training.

9. Are we allowed to use test sets in training? 

Additional manual annotations on our testing data are strictly prohibited. We also do not encourage the use of testing data in any way during training, with or without labels, because the task is supposed to be fairly evaluated in real life where we don’t have access to testing data at all. Although it is permitted to perform algorithms like clustering to automatically generate pseudo labels on the testing data, we will choose a winning method without using such techniques when multiple teams have similar performance (~1%). Finally, please keep in mind that, like all the previous editions of the AI City Challenge, all the winning methods and runners-up will be requested to submit their code for verification purposes. Their performance needs to be reproducible using the training/validation/synthetic data only.

10. Are we allowed to use data/pre-trained models from the previous edition(s) of the AI City Challenge?

Data from previous edition(s) of the AI City Challenge are allowed to be used.

11. Do the winning teams and runners-up need to submit papers and present at the workshop?

Track 1 – Multi-Camera 3D Perception (Sim2Real)

1. Is calibration available for each camera?

The comprehensive camera calibration information is available for each camera, including 3-by-4 camera matrix, intrinsic parameters, extrinsic parameters, etc.

2. What is the standard of labeling visible 2D bounding boxes?

The annotations of the test set are generated based on the same standards as the training and validation set.

• • For occluded objects (objects that are blocked by an object within the camera frame), objects must satisfy both the visibility in height and width requirements.
  • For objects that are truncated – objects that are cut off via the camera frame, the objects must satisfy EITHER of the conditions in visibility for height OR the visibility for width.
  • Here are the definitions for visibility in height and width:
    • Visibility for height
      • If the head is visible and 20% of the height is visible then, label the object.
      • If the head is not visible, then label the object if 60% of the height is visible.
    • Visibility for width
      • More than 60% body width visible should be labeled.

3. How are the object IDs used for evaluation? Do the submitted IDs need to be consistent with the ground truths?

We use the HOTA metric for evaluation. The IDs in the submitted results do not need to match the exact IDs in the ground truths. We will use bipartite matching for their comparison, which will be based on IoU of 3D bounding boxes in the global coordinate system.

Track 2 – Transportation Safety Understanding and Captioning (Sim2Real)

1. Can we use any pre-trained models from earlier AI City Challenge versions of this track?

No. Models pre-trained or fine-tuned on AI City Challenge data from previous years or data from other tracks from this year’s challenge are not allowed for this track.

2. Can other pre-trained models be used for this challenge track?

Yes. Teams may fine-tune general pre-trained models with open weights such as Qwen 3.6 or Gemma 4. As solutions to this challenge should be reproducible, teams should refrain from using paid API-based models such as Google Gemini or OpenAI GPT.

3. Can we modify the synthetic videos to increase their realism before we train our models?

Yes. Teams may use generative models that may enhance the realism of the scenes, such as the NVIDIA Cosmos line of models, which may better align the synthetic data and the real data distributions.

Track 3 – Anomalous Events in Transportation

1. Are closed-source / paid API-based models (e.g., OpenAI GPT, Google Gemini, Anthropic Claude) allowed in Track 3?

Yes. Track 3 permits publicly accessible models or services, including paid API-based models, to be used in any role in your pipeline:

• • generating, refining, augmenting, or cleaning training data;
  • producing intermediate outputs (captions, structured event descriptions, reasoning chains) consumed by a downstream model;
  • serving as the final inference system that produces submitted predictions.

This applies uniformly to all ten in-domain task types and to both out-of-domain evaluations.

The only hard constraint is that any external models or data you use must be publicly accessible to all participants, such as through a public API, open download link, public repository, or official project page with access instructions. Teams may also use their own internally developed models in any of the above roles, provided they are trained only on permitted challenge data and publicly accessible external data/models. No private data may be used in their training. Internally developed models do not need to be published prior to award verification. Award candidates will be required to open-source the components they train and control (code, model weights, training recipes, and any custom inference-pipeline code) by the open-source deadline announced on the official challenge website. Components that participating teams did not train or control (e.g., closed-source APIs accessed via their public service) are not required to be released.

Because closed-source services are not bit-for-bit reproducible, award candidates must clearly disclose the service, model version, role in the pipeline, and prompts/configuration for each closed-source API component used. Code verification will focus on confirming that no private data or non-public model was used, rather than on bit-exact replay of closed-source API calls.

2. Must we train a single end-to-end VLM, or are modular and agentic systems allowed?

Track 3 requires a unified system: one inference pipeline that handles all task types in the in-domain test set and both out-of-domain evaluations. Given a video and a task-specific prompt, the system must produce an answer through that unified pipeline. Task-specific prompts, parsing, and routing inside the pipeline are allowed, but teams may not submit separately tuned or specialized systems per task type or per test set.

The internal architecture of the unified system is up to you, and may be:

• • a single fine-tuned end-to-end VLM model,
  • an agentic pipeline (planning, tool use, multiple model calls in sequence),
  • a modular multi-stage system in which different components play different roles (e.g., a perception/captioning module feeding a separate reasoning module).

Any external models or data you use must be publicly accessible to all participants, such as through a public API, open download link, public repository, or official project page with access instructions. Teams may also use their own internally developed models, provided they are trained only on permitted challenge data and publicly accessible external data/models. No private data may be used in their training. Internally developed models do not need to be published prior to award verification. Award candidates will be required to open-source the components they train and control (code, model weights, training recipes, and any custom inference-pipeline code) by the open-source deadline announced on the official challenge website. Each submission must clearly categorize the developed solution.

3. Are we allowed to use additional public datasets, web-sourced videos, or other external data?

Yes. Track 3 allows publicly accessible data to be added to your training mixture, including additional public video datasets, web-sourced videos under permissive licenses, synthetic data from public generative models, and distilled outputs from public or commercial APIs. External data sources used must be declared in the technical report. Private data, scraped data without a clear public license, and non-public datasets are not allowed.

4. Can we modify, rewrite, or extend the released training annotations?

Yes. The released annotations may be used as-is, post-processed (for example, by rewriting chain-of-thought reasoning, filtering noisy items, or generating additional Q&A pairs), or combined with annotations you generate yourself using public or commercial models. Any additional generated annotations or substantial annotation post-processing should be described in the technical report.

Track 4 – Text-Based Person Re-Identification (Sim2Real)