Paper Overview
- Field: Computer Vision
- Authors: Mobasshir Bhuiya Shagor, Nilambar Halder Tonmoy, Md Saef Ullah Miah
- Published: 2026-11-10
- arXiv: 2611.14476
- Image-random and pooled cattle-disjoint splits over the same development cohorts
- A separate future-2025 evaluation
- A locked future-2026 evaluation
- The gap between image-random and cattle-disjoint splits is task-dependent, not a universal leakage penalty.
- Future-cohort behavior is attribute- and cohort-specific: breed degrades relative to pooled evaluation; mean weight and age fail to beat a constant baseline in 2025; height remains above baseline; coat color transfers well in 2025 but degrades in 2026.
- Multi-task training improves all five point estimates in single-seed ablations.
- Static-view aggregation helps selectively; zero-shot video-frame aggregation exposes a static-to-video domain gap.
Summary
Computer vision offers a scalable, non-contact means of estimating livestock traits, but correlated images and video frames make reliable evaluation difficult. Image-random splitting can place views of the same animal in both training and test sets, while deployment further requires transfer to animals captured in future years.
The authors introduce BoviShift, a corpus reconstructed from six consecutive annual cohorts of an operational cattle marketplace collected from 2021 to 2026. It contains 2,657 cattle, 10,610 still images, and 2,566 walk-around videos across nine recorded breed categories.
Evaluation Setup
Three splitting strategies are evaluated through four configurations:
A fixed ResNet-50 multi-task baseline predicts weight, age, height, breed, and coat color.
Key Findings
--- *Auto-collected on 2026-11-12*