Our laboratory participated in the 10th Affective & Behavior Analysis in-the-Wild (ABAW) Competition, held in conjunction with the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2026). The ABAW Workshop took place on June 3, 2026, in Denver, USA. The competition focused on the analysis of human affect and behavior in unconstrained real-world conditions and included six challenges. The RAS team from our laboratory participated in the Valence-Arousal Estimation Challenge on continuous human affect estimation and took first place.

As part of the competition, the following paper was published:

Ryumina E., Markitantov M., Axyonov A., Ryumin D., Dolgushin M., Dresvyanskiy D., Karpov A. From Faces to Behavior Analysis: Adaptive Multimodal Fusion for Valence-Arousal Estimation in-the-Wild // Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 10th ABAW Workshop, IEEE, Denver, USA, 2026, pp. 5419–5428.

Continuous emotion recognition in the Valence-Arousal space under real-world conditions remains a challenging problem due to large variations in appearance, head pose, illumination, facial occlusions, and individual patterns of affective expression. The paper presents a novel multimodal approach combining three complementary sources of information: face, behavior, and audio. The face modality uses GRADA-based frame-level features and Transformer-based temporal modeling. Behavior-related information is extracted from video segments using Qwen3-VL-4B-Instruct, while its temporal dynamics are modeled with the Mamba architecture. The audio modality is based on WavLM-Large with attention-statistics pooling and includes cross-modal filtering to reduce the influence of unreliable and non-speech segments. Two adaptive modality-fusion strategies are investigated: Directed Cross-Modal Mixture-of-Experts and Reliability-Aware Audio-Visual Fusion. Experiments are conducted on the audiovisual Aff-Wild2 dataset following the protocol of the 10th ABAW Valence-Arousal Estimation Challenge. The best multimodal approach achieves an average Concordance Correlation Coefficient of CCC=0.658 on the development set and CCC=0.622 on the test set. As a result of the competition, Team RAS took first place in the Valence-Arousal Estimation Challenge.