Abstract
Tracking algorithms based on computer vision provide the location of each individual at the frame level. This information can be used for social network analysis by calculating proximity between individuals. However, down sampling frames is necessary for efficient computation and long-period analysis. Down sampling can be achieved by increasing the interval between frames (sampling interval). This study investigated how different sampling intervals impact proximity networks. Video data were collected over three consecutive
days in 21 pens, each containing six pigs. A tracking-by-detection method was applied to detect the location and track each individual by using bounding boxes. Networks were constructed with different proximity definitions (based on the surface of overlap and the distance between pairs of bounding boxes). For each proximity definition, a complete network (2s sampling interval) and reduced networks (sampling interval from 10s up to an hour) were built. We explored how different sampling intervals affected the correlation between the metric of the reduced and the complete networks. The metrics examined were degree, eigenvector centrality, clustering coefficient, and radius. Our results suggest that
for the proximity definitions based on the overlap of the bounding boxes, larger sampling intervals (e.g. 15min) can provide similar information as shorter ones (correlations >0.90), whereas for proximity definitions based on distance, the interval should not go below 6min. Furthermore, clustering coefficient is more sensitive to down sampling than degree, eigenvector centrality, or radius. Our findings demonstrate that a large sampling interval impacts the network, potentially influencing the conclusions that are drawn, and provide a method to select the optimal sampling interval.
days in 21 pens, each containing six pigs. A tracking-by-detection method was applied to detect the location and track each individual by using bounding boxes. Networks were constructed with different proximity definitions (based on the surface of overlap and the distance between pairs of bounding boxes). For each proximity definition, a complete network (2s sampling interval) and reduced networks (sampling interval from 10s up to an hour) were built. We explored how different sampling intervals affected the correlation between the metric of the reduced and the complete networks. The metrics examined were degree, eigenvector centrality, clustering coefficient, and radius. Our results suggest that
for the proximity definitions based on the overlap of the bounding boxes, larger sampling intervals (e.g. 15min) can provide similar information as shorter ones (correlations >0.90), whereas for proximity definitions based on distance, the interval should not go below 6min. Furthermore, clustering coefficient is more sensitive to down sampling than degree, eigenvector centrality, or radius. Our findings demonstrate that a large sampling interval impacts the network, potentially influencing the conclusions that are drawn, and provide a method to select the optimal sampling interval.
Original language | English |
---|---|
Title of host publication | 11th European Conference on Precision Livestock Farming (ECPLF 2024) |
Editors | D. Berckmans, P. Tassinari, D. Torreggiani |
Publisher | European Association for Precision Livestock Farming |
Pages | 1737-1744 |
ISBN (Electronic) | 9791221067361 |
ISBN (Print) | 9798331303549 |
Publication status | Published - Oct 2024 |
Event | 11th European Conference on Precision Livestock Farming - Bologna, Italy Duration: 9 Sept 2024 → 12 Sept 2024 |
Conference/symposium
Conference/symposium | 11th European Conference on Precision Livestock Farming |
---|---|
Country/Territory | Italy |
City | Bologna |
Period | 9/09/24 → 12/09/24 |
Keywords
- proximity
- social network
- tracking data
- computer vision