File Instruction

The instructions for downloading VFP290K is listed as follows: 

1.  We provide the list of videos and 4 JSON files (Light conditions, Camera heights, Background, and benchmark).

2.  Each JSON file has training, validation, and test videos for reproducibility.

METADATA Information

Our metadata for the VFP290K dataset consists of 7 definitions, including light conditions, camera heights, background, location, view,  scene, occlusion, and the number of frames:

Light conditions (M_day, night). Light conditions are one of the most important features impacting the fallen person detection performance, as the quality of the image is highly dependent on it. We consider two light conditions, M_day and M_night. In VFP290K dataset, it is represented 0 for M_day and 1 for M_night. 

Camera heights (M_low, high). Camera height is directly related to model CCTV and indoor camera environments. It is not trivial to model all the possible height requirements for different CCTV devices, as there are many different types of CCTV with varying positions. To address this issue, we film the video into several camera heights with M_low and M_high, where M_low is a camera height about 1 ~ 3m and M_high represents higher than 3m.

Background (M_street, park, building).  One of the main contributions of our approach is to have numerous backgrounds, from public areas to indoor environments. We divide the background into the following three sub-categories: street, park, and building. We assign 0, 1, 2 for the street, park, and building backgrounds, respectively.

Location (M_location) & View (M_view). We identify each location in the background by enumerating the videos. The total number of locations for each background is 13, 30, and 6 for a street, a park, and a building, respectively.  Also, we report each view in the location by enumerating the videos. The minimum, average, and a maximum number of the viewpoints are 1, 2, 6, and 15, respectively. We try to differ the viewpoints in a certain area, such as a playground.

Scene (M_Scene).  Along with the three background categories, we also categorize the filming locations as scenes. We enumerate each scene by each location divided by each view, composing 131 different scenes in detail. We record each scene by varying the filming view dramatically in each location.

Occlusion (M_occlusion).  We specify whether the video contains occlusion or not, assigning 0 for the video without occlusion and 1 for the video with occlusion.

Number of Frames (M_Scene). We specify the number of frames for each video to control the distribution of the benchmark.

Release history

• Our VFP290K dataset has a digital object identifier (DOI) as follows: https://doi.org/10.23056/VFP300K_DASHLAB

• Our Github link for VFP290K benchmark experiments as follows: https://github.com/DASH-Lab/VFP290K

• Version 0.1

The total files is about 190 GB

• Download link for VFP290K : Email "swoo@g.skku.edu" for the request including your affliation, purpose, etc.