Answers are grounded in this paper and cite their sources.
We compared the stepwise Unrest Index with tracker-derived travel distance and dense optical flow in eight overhead recordings of groups of ten Ross broilers recorded 15–40 days after housing. The principal recording lasted 29.8 min and the other seven about 60 min; four supplied no frames to model training or validation. In each recording, one tracked pass of an instance-segmentation model gave raw box centers, mask-area centroids and tracker box centers for identical instances. Agreement was Spearman’s ρ with paired block bootstrap intervals. With instances held fixed, the mask centroid showed no consistent difference from the box center in agreement with optical flow (Δρ −0.048 to +0.051). The higher agreement previously attributed to mask centroids was reproduced only when the model changed (Δρ +0.042 to +0.120; all within-recording intervals above zero). The 95th-percentile Hausdorff distance gave small gains over the maximum (Δρ +0.020 to +0.031 against flow), which do not identify an optimal percentile. Agreement with travel distance (ρ = 0.77–0.89) was stronger than with optical flow (ρ = 0.28–0.59), and switching the travel-distance coordinates from mask centroids to raw box centers reduced the correlation by 0.15–0.23. Restricting flow to the pen floor raised index–flow agreement. Distances are in pixels, and all results concern about ten birds per frame.
No takes yet. Share an insight, caveat, or question.
Hubner et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: