3rd Ambivalence / Hesitancy (AH) Video Recognition Challenge

01 · OverviewAbout the challenge

The Ambivalence/Hesitancy (AH) Video Recognition Challenge invites teams to build models that decide whether a person is ambivalent or hesitant in a given video. Now in its third edition, the challenge is organized within the 11th Workshop and Competition on Affective & Behavior Analysis in-the-Wild (ABAW) at ECCV 2026, and is powered by the richly annotated BAH dataset.

Upon registration, teams are granted full access to the BAH training data and can develop, train and benchmark their own methods, before being evaluated on a held-out private test set. The top teams are invited to publish their approach in the ECCV 2026 workshop proceedings.

02 · BackgroundWhat is ambivalence & hesitancy?

Ambivalence and hesitancy (A/H) are closely related constructs and a primary reason why people delay, avoid, or abandon health-behaviour changes. A/H is a subtle, conflicting emotion that places a person between positive and negative orientations, or between acceptance and refusal to act. It is a barrier to initiating behaviour change and a trigger for discontinuing interventions, which makes detecting it valuable for digital, personalized behaviour-change interventions.

To detect A/H, experts rely on conflicting emotional cues, positive and negative, expressed through facial, language, audio and body-language modalities. Conflicts across modalities (cross-modality) are comparatively easy to spot, but a conflict within a single modality can also signal A/H and is far harder to detect. Our work [1] showed that standard deep multimodal models and fusion techniques achieve only low accuracy on this task, indicating they are not yet well equipped for A/H recognition.

Can a model tell, from a short video, when someone is of two minds, caught between acceptance and refusal?

03 · The taskVideo-level A/H recognition

This edition keeps the same task as the previous one: video-level binary prediction. Given a video, predict whether it contains ambivalence/hesitancy: 1 for presence of A/H, 0 for absence. Since A and H manifest similarly in practice, BAH provides a single binary annotation covering both, without distinguishing between them.

Teams are free to explore a wide range of learning setups, including:

• Supervised and self-supervised learning
• Domain adaptation, test-time adaptation and personalization (BAH is participant-based)
• Zero-shot and few-shot learning
• Standard multimodal models, vision-language models (VLMs) and multimodal LLMs with parameter-efficient fine-tuning (PEFT)
• Temporal modeling, multimodal alignment and specialized fusion to capture conflicting affect within and across modalities

Interpretable solutions are encouraged, for example highlighting when A/H occurs in a video, or which modalities, cues or conflicts drive a prediction. The rich frame- and video-level annotations of BAH can be used for both training and evaluation.

04 · DatasetThe BAH video dataset

On registration, teams receive a fully annotated (video- and frame-level) version of the BAH dataset [1], collected for multimodal recognition of A/H in videos. It contains 1,427 videos totalling 10.60 hours, captured from 300 participants across Canada who answered a predefined set of questions designed to elicit A/H, mirroring real-world online behaviour-change interventions.

• Expert annotations with timestamps marking where A/H occurs, plus frame- and video-level labels with A/H cues
• Multimodal data: raw videos, cropped and aligned faces per frame, audio, and speech-to-text transcripts with timestamps
• Annotator cues, participant metadata, and predefined participant-wise splits (training, validation, test)
• Up to seven videos per participant; the data is divided participant-wise to prevent identity leakage between splits

Teams train on the BAH training set under any form of supervision and report performance on the public test set using the provided bah_metrics.py. A second, unlabeled private test set is released near the end of the challenge and is used for the official ranking. For full details, figures and benchmarks, see the BAH dataset page.

05 · RulesWhat's allowed

Teams may use any publicly available or private pre-trained model and any public or private dataset containing any type of annotation (e.g. valence/arousal, basic or compound emotions, action units). Other ambivalence/hesitancy datasets, where available, may be used in addition to BAH. Any external dataset used must be disclosed in the paper. Teams are encouraged to develop solutions specifically tailored to A/H recognition rather than relying on generic emotion models.

06 · EvaluationHow submissions are scored

The ranking metric P is the macro-averaged F1 score (Macro F1) at the video level across both classes, presence (1) and absence (0) of A/H, computed over the private test set. The average precision (AP) of the positive class is also reported.

During the one-week test period, teams submit per-video predictions on the private test set by email to the organizers. Up to 5 trials are allowed, either all at once or one at a time, the latter lets organizers return per-trial feedback so teams can adjust their approach. The best trial is used to rank teams and determine the winners. The exact submission format is communicated on the test-release date.

07 · BaselineStarting point

A baseline performance of P = 0.2827 was obtained on the BAH public test set using a zero-shot multimodal LLM (Video-LLaVA) with a simple prompt and the vision modality only. Teams can also build on a standard multimodal model that leverages vision, audio and text, adapting it from frame-level to video-level prediction. Both the baseline code and the BAH dataset code are openly available as references.

08 · EditionsA three-edition history

• 1st edition — ABAW 8th @ CVPR 2025: frame-level prediction of A/H, which attracted several teams.
• 2nd edition — ABAW 10th @ CVPR 2026: moved to video-level prediction, is there A/H in a given video? Solutions are summarized on arXiv.
• 3rd edition — ABAW 11th @ ECCV 2026: the same video-level task, now with access to the larger, richly annotated BAH dataset.

09 · TimelineKey dates

10 · AwardsPublication & recognition

The top 3 winning teams will contribute paper(s) describing their approach, methodology and results, and are required to make their code public. Accepted papers will be part of the ECCV 2026 workshop proceedings.

11 · RegisterHow to take part

All participating teams must register. Registration involves completing a form and signing an EULA. Both the form and the EULA must be completed and signed by a person holding a full-time faculty position at a university, higher-education institution, or equivalent organization, the signee cannot be a student (undergraduate, postgraduate, Ph.D. or postdoctoral). Once the signed form is submitted, the organizers will share access details for the BAH video dataset.

Find the registration form, EULA and the full call on the ABAW 11th website, and reach the organizers at livia-datasets@livia.etsmtl.ca for questions about registration or data access.

12 · TeamOrganizers

Eric Granger, Alessandro Lameiras Koerich and Marco Pedersoli (LIVIA, ÉTS Montréal), and Simon L. Bacon (Montreal Behavioural Medicine Centre, Concordia University). Contact: livia-datasets@livia.etsmtl.ca.

13 · ReferenceCite the BAH dataset

The challenge is built on the BAH dataset, see the full BAH dataset page for details, figures and benchmarks.

@article{gonzalez-25-bah,
  title={BAH Dataset for Ambivalence/Hesitancy Recognition in
         Videos for Digital Behavioural Change},
  author={González-González, M. and Belharbi, S. and Zeeshan, M. O.
          and Sharafi, M. and Aslam, M. H. and Pedersoli, M. and
          Koerich, A. L. and Bacon, S. L. and Granger, E.},
  journal={arXiv preprint arXiv:2505.19328},
  year={2025}
}