Context The interpretation of human interactions in images or videos has significantly improved with the emergence of Large Language Models (LLMs) and Vision-Language Models (VLMs). However, these large models, whether used directly or distilled into specialized models, still have significant limitations, particularly in accurately attributing interactions to the correct person in dense scenes and discriminating actions in the presence of objects.   Evaluation protocols and databases for this task do not always accurately reflect the true capabilities of the methods due to issues such as annotation imprecision or overly rigid semantic metrics. What do we expect from you? To address these problems, the internship will focus on the following objectives: - Conduct a state-of-the-art review of existing databases and analyze their biases (e.g., precision of detection boxes). - Propose a semi-automatic pipeline for correcting these biases. - Identify the biases and gaps in the metrics commonly used in the state-of-the-art. - Propose a new benchmark, addressing various application domains. - Evaluate the main state-of-the-art approaches on this benchmark. - Write a publication about this benchmark. #Cea List