DATASETS
1) Abnormal Human Activity Dataset
The Abnormal Human Action dataset (AbHA) Abnormal Human Action Dataset (AbHA) includes five commonly occurring abnormal actions in day-to-day lives of the elderly people: ‘chest pain’, ‘headache’, ‘fainting’ and ‘falling backward’ and ‘falling forward’. Each action is performed by eight different individuals and repeated two times. Hence, total generated samples are 80 (8×2×5). The dataset is captured by Microsoft’s Kinect depth sensor v1. The resolution of depth videos is 640X480. 20 skeleton joints are captured in 3D coordinate system ('action_1mainWorld.txt') and depth coordinate system (‘action_1mainM.txt’) for each frame.
To download this dataset, you have to fill the form https://forms.gle/9FA4XuvNU7ugv1gn8
2) EmoGDB dataset
To further enhance the research in the area of movie genre detection, especially for Indian cinema, we developed an EmoGDB (Emotion-based Genre Detection for Bollywood) dataset, which is specifically related to detecting the genre of Hindi Bollywood movies. The major work which has contributed to study the relationship between the field of psychology and cinematography. We adopt six emotion categories, namely: Happy, Surprise, Anger, Sad, Fear, and Neutral. These are the major emotions that are evoked while watching any movie. The prime advantage of this dataset lies in the fact that it is labeled with five-movie genres along with the six different types of emotions (which are elicited while watching a movie trailer) corresponding to each genre. To the best of our knowledge, no dataset has been developed in the literature that provides such rigorous affective-based information for different movie genres. The trailers have a broad range of release dates (1996 to 2020). Fig. 1 shows some sample images from our dataset (movie name: genre). As seen in the figure, each genre is provoking a wide range of emotions.
Fig. 1. Sample images from EmoGDB dataset (a) 1920: Horror (b) Ae Dil Hai Mushkil : Romance (c) Gameover : Thriller (d) Behen Hogi Teri : Comedy (e) Chhichhore : Drama (f) Baahubali 2_The Conclusion : Action.
We browsed and collected a list of famous Bollywood movies of different genres from four popular film libraries: IMDB, NetFlix, Hotstar, and Amazon Prime. We only focused on those movie trailers which have a common genre on each of these libraries. Our dataset is created and structured to allow the research community to use it with ease. We created one folder per movie. The file structure of the dataset per movie is illustrated in Fig. 2.
Each movie folder contains two sub-folders of uncropped facial frames and cropped facial frames, along with the corresponding movie trailer. The naming format of each folder and sub-folder is shown in Fig. 3. The cropped frame information is stored in the CSV file together with the output labels. The CSV contains the following information: Frame_Name, Movie_Name, Genre, and Emotion. The final trailers belong to the following genres, namely: Action (17), Comedy (16), Drama (17), Horror (17), Romance (16), and Thriller (17). The reason to limit our work to these genres is that mostly all the movies can roughly be classified into at least one of these genres. Secondly, we propose that these genres elicit strong induced emotions, which can be crucial for classifying these movie trailers into multiple genres. EmoGDB dataset consists of roughly over 1,00,000 frames corresponding to 100 Bollywood movie trailers. Since the length of a Bollywood film is very long (up to 2-3 h), hence we concentrated on extracting the features from the movie trailer, which typically has a duration of around 2–4 min.
To download this dataset https://forms.gle/QPVNk67UhM9HxTEP9