Real-World Pig Weight Estimation via Morphometry-Guided Coupled Learning from Top-View Depth Images

Non-contact pig weight estimation is essential for precision livestock farming, yet most vision-based methods model body weight as a direct black-box regression problem, neglecting the explicit role of animal morphometry. In practice, pig body weight is largely determined by fundamental structural traits, particularly body length and body width, which are also the primary indicators used in manual on-farm measurements. Motivated by this observation, we propose a morphometry-guided coupled learning framework that explicitly introduces body length and body width as intermediate variables for weight estimation from a single top-view depth image. Unlike conventional multitask learning that relies on implicit feature sharing, our framework explicitly leverages the specific geometric constraints of the walking channel, where the pig's movement is parallel to its length and perpendicular to its width. Rather than using standard global attention regression, we engineer an architecture with two branches equipped with novel directional (horizontal and vertical) attention modules. This design spatially decouples the visual information, compelling the network to extract highly discriminative representations of actual body dimensions independently before fusion. Validation on a real world dataset comprising 315 pigs (10-150 kg) yields a mean absolute error (MAE) of 3.672 kg. Outperforming traditional multitask methods, this result effectively demonstrates that explicitly decoupling spatial features through geometric priors provides a mathematically and practically superior alternative to conventional multitask regression.