Researchers collected about 127 audio-visual dog recordings in Mangalore to build a multimodal canine emotion dataset covering Happy, Sad, Angry, and Relaxed

It constructs a real-world canine emotion audio-visual dataset containing approximately 127 recordings across four emotion classes: Happy, Sad, Angry, and Relaxed. Compared with prior work, the dataset provides both audio and video modalities and is explicitly aimed at canine emotion recognition rather than behavior alone or a single modality. The text states the data were gathered from various localities of Mangalore, with approximately 127 recordings and four emotion classes; it does not provide per-class sample counts, number of dogs, recording duration, or device details.

Emotion labels were identified and assigned through the application of various clustering algorithms rather than relying entirely on manual annotation. This pipeline introduces unsupervised clustering into canine emotion label generation, differing from purely manual labeling or observation-only grouping. The text mentions the use of various clustering algorithms for emotion identification and label assignment, but does not name the specific algorithms, clustering evaluation metrics, or label consistency validation results.

The dataset is suitable for multimodal, audio-based, video-based, and emotion classification systems, and can support analysis of correlations between dog behavior, expression, and emotions. It offers a training and evaluation basis for automatic canine emotion recognition algorithms and supports cross-modal correlation analysis. The text is primarily a use-case statement and reports no baseline models, classification accuracy, or cross-modal comparison results.

This dataset is intended for researchers and developers studying automatic canine emotion recognition. It applies to training and evaluation of multimodal, audio-based, video-based, and emotion classification systems, and can also be used to analyze correlations between dog behavior, expression, and emotions. Its designed setting is real-world recording in various localities of Mangalore, and the text positions it as a foundational resource for investigating the feasibility of recognizing dog emotions from audio-visual data.

The text is an incomplete read, missing the file list, data tables, figures, and experimental sections, so it cannot confirm per-class sample counts, number of individual dogs, total recording duration, specific clustering algorithm types and parameters, label reliability validation methods, or whether baseline classification results exist. These gaps affect a complete judgment of dataset scale, class balance, and usability, and are open questions requiring consultation of the original text.