The challenge can be found in CodaLab here: https://competitions.codalab.org/competitions/31432#learn_the_details-overview

• Overview
  
  Streaming platforms are pervasive nowadays, and everyone has access to them. This raises major concerns for minors who are exposed to content that may not be appropriate for their age. Automated tools for tagging online content according to the presence of questionable material can have a great impact in protecting users from being exposed to disturbing and inappropriate material. With this goal in mind, we are organizing a grand challenge collocated with ICMI 2021 on labeling questionable content in the wild.
  
  You are invited to join this grand challenge, where we approach the task of the automated detection of questionable comic mischief in videos. By comic mischief, we mean the appearance of mild harm inflicted on video characters in a manner that is intended to be humorous or funny. The content itself can be present in more than one single modality (soundtrack, dialogue, or video). We intend to motivate new multimodal research that can detect the presence of this type of content. Among the primary applications is the automated labeling of comic mischief in videos easily accessible to young viewers. Participants of the challenge are asked to develop solutions for automatically recognizing comic mischief categories from videos.
  
  
  • Challenge
    
    Participants of this challenge should develop methods for predicting the comic mischief categories from audiovisual information. Depending on the track (see below) participants will provide binary (comic-mischief vs. no-comic-mischief content) or multiclass (mature humor, slapstick humor, gory humor, sarcasm or none) labels for each scene. Please note that a single scene may be associated to more than a single comic-mischief label, that is, this is a multi label multi class classification problem.
    
    A description of the considered categories is provided for your reference:
    
    
    • Mature humor: Comic content with adult themes, such as content including sexual innuendos, political satire, or other themes that are not easily grasped by children.
      
    
    
    • Slapstick humor: Specific style of humor that depicts intentional or accidental violence, usually involving props (such as a slapstick)..
      
      
    • Gory humor: This category refers to content that presents graphic violence, or the result of extremely violent actions, in a humoristic manner.
      
      
    • Sarcasm: Content that relies on irony as the primary mechanism for delivering the comic content.
      
      
    • Subtasks
      
      The MOCHA challenge comprises two subtasks that aim at evaluating different settings of the problem:
      
      
      • Track 1: Binary detection of comic mischief at the scene level - Participants should assign a scene a binary label indicating the presence of comic mischief. The idea is to provide a localized labeling of content with coarse (present/absent) binary labels. Task participants will be provided with videos already segmented by scenes.
        
      
      
      • Track 2: Fine-grained detection of comic mischief at the scene level - The difference between this task and Task 1 above is that instead of a binary labeling scheme, we will use the four fine-grained labels defined above: mature humor, slapstick humor, gory humor, sarcasm, plus none.
        
        
      Submissions: Participants should prepare a prediction file similar to the one provided as ground-truth training data for each of the tracks (see the files binary.sol and multilabel.sol). The file should be zipped (please do not create / include any folder).
      
      Evaluation: For system submissions, we will rank systems based on F2-measure. Considering that this research’s ultimate goal is to detect questionable content, we propose to use F2-measure as the ranking metric for system submissions. In this task, it is more important to minimize false negatives, and thus the emphasis on setting the β parameter of the F-measure to 2. Other evaluation metrics like F1, AUC, and Accuracy could analyze results and solve ties, but the leading evaluation measure will be F2-measure.