Automatically understanding funny moments (i.e., the moments that make people
laugh) when watching comedy is challenging, as they relate to various features,
such as body language, dialogues and culture. In this paper, we propose
FunnyNet-W, a model that relies on cross- and self-attention for visual, audio
and text data to predict funny moments in videos. Unlike most methods that rely
on ground truth data in the form of subtitles, in this work we exploit
modalities that come naturally with videos: (a) video frames as they contain
visual information indispensable for scene understanding, (b) audio as it
contains higher-level cues associated with funny moments, such as intonation,
pitch and pauses and (c) text automatically extracted with a speech-to-text
model as it can provide rich information when processed by a Large Language
Model. To acquire labels for training, we propose an unsupervised approach that
spots and labels funny audio moments. We provide experiments on five datasets:
the sitcoms TBBT, MHD, MUStARD, Friends, and the TED talk UR-Funny. Extensive
experiments and analysis show that FunnyNet-W successfully exploits visual,
auditory and textual cues to identify funny moments, while our findings reveal
FunnyNet-W’s ability to predict funny moments in the wild. FunnyNet-W sets the
new state of the art for funny moment detection with multimodal cues on all
datasets with and without using ground truth information.

FunnyNet-W: Exploiting Multimodal Cues for Funny Moment Detection in Videos

Understanding humor and what makes people laugh is a complex task that involves several factors, including body language, dialogues, and cultural references. In the field of multimedia information systems, detecting funny moments in videos has been a challenge due to the multi-disciplinary nature of the concept. However, a recent paper introduces a groundbreaking model called FunnyNet-W, which leverages cross- and self-attention mechanisms to predict funny moments using visual, audio, and text data.

The unique aspect of FunnyNet-W is its reliance on modalities naturally present in videos, rather than relying on ground truth data like subtitles. The model utilizes video frames to capture visual information critical for scene understanding. Additionally, it leverages audio cues associated with funny moments, such as intonation, pitch, and pauses. Text data extracted using a speech-to-text model is also processed by a Large Language Model to extract valuable information. By combining these modalities, FunnyNet-W aims to accurately identify and predict funny moments in videos.

The paper also introduces an unsupervised approach for acquiring labels to train FunnyNet-W. This approach involves spotting and labeling funny audio moments. By doing so, the model can learn from real-life instances of humor rather than relying solely on pre-defined annotations.

To evaluate the performance of FunnyNet-W, the researchers conducted experiments on five datasets, including popular sitcoms like The Big Bang Theory (TBBT), Modern Family (MHD), and Friends, as well as the TED talk dataset UR-Funny. The comprehensive experiments and analysis showed that FunnyNet-W successfully utilizes visual, auditory, and textual cues to identify funny moments. Moreover, the results demonstrate FunnyNet-W’s ability to predict funny moments in diverse and uncontrolled video environments.

From a broader perspective, this research contributes to the field of multimedia information systems by showcasing the effectiveness of combining multiple modalities for humor detection. FunnyNet-W’s reliance on visual, audio, and textual data highlights the multi-disciplinary nature of understanding funny moments in videos. By incorporating insights from computer vision, audio processing, and natural language processing, this model represents a step forward in multimodal analysis.

Furthermore, the concepts presented in FunnyNet-W have implications beyond just humor detection. The model’s ability to leverage multiple modalities opens up possibilities for applications in various domains. For example, this approach could be utilized in animations to automatically identify comedic moments and enhance the viewer’s experience. Additionally, the integration of visual, audio, and textual cues can also be valuable for improving virtual reality and augmented reality systems, where realistic and immersive experiences rely on multimodal input.

In conclusion, FunnyNet-W establishes a new state-of-the-art for funny moment detection by effectively exploiting multimodal cues across various datasets. This research not only advances our understanding of humor detection but also demonstrates the power of combining visual, audio, and textual information in the wider context of multimedia information systems, animations, artificial reality, augmented reality, and virtual realities.

Read the original article