Can 3D point cloud data improve automated body condition score prediction in dairy cattle?

Body condition score (BCS) is a widely used indicator of body energy status and is closely associated with metabolic status, reproductive performance, and health in dairy cattle; however, conventional visual scoring is subjective and labor-intensive. Computer vision approaches have been applied to BCS prediction, with depth images widely used because they capture geometric information independent of coat color and texture. More recently, three-dimensional point cloud data have attracted increasing interest due to their ability to represent richer geometric characteristics of animal morphology, but direct head-to-head comparisons with depth image-based approaches remain limited. In this study, we compared top-view depth image and point cloud data for BCS prediction under four settings: 1) unsegmented raw data, 2) segmented full-body data, 3) segmented hindquarter data, and 4) handcrafted feature data. Prediction models were evaluated using data from 1,020 dairy cows collected on a commercial farm, with cow-level cross-validation to prevent data leakage. Depth image-based models consistently achieved higher accuracy than point cloud-based models when unsegmented raw data and segmented full-body data were used, whereas comparable performance was observed when segmented hindquarter data were used. Both depth image and point cloud approaches showed reduced accuracy when handcrafted feature data were employed compared with the other settings. Overall, point cloud-based predictions were more sensitive to noise and model architecture than depth image-based predictions. Taken together, these results indicate that three-dimensional point clouds do not provide a consistent advantage over depth images for BCS prediction in dairy cattle under the evaluated conditions.

💡 Research Summary

This study provides a comprehensive head‑to‑head comparison of depth‑image and three‑dimensional (3D) point‑cloud representations for automated prediction of body condition score (BCS) in dairy cattle. BCS is a key indicator of energy balance, health, and reproductive performance, yet conventional visual scoring is subjective, labor‑intensive, and prone to inter‑observer variability. While depth sensors have become popular because they capture geometric information independent of coat color, recent advances in 3D point‑cloud reconstruction promise richer morphological detail. However, direct comparisons between the two modalities have been scarce.

Data were collected on a commercial dairy farm in central Georgia, USA, using an Intel RealSense D455 depth camera installed above a chute at the milking parlor exit. Over two periods (November 2024 and February 2025) a total of 1,020 cows were imaged, yielding 10,228 depth CSV files and corresponding RGB images. Each cow was scored by two trained evaluators on a 5‑point BCS scale with 0.25‑point increments; only cows with exact agreement were retained. The dataset thus spans ages 3–10 years and BCS values from 2.0 to 4.25.

The raw depth rasters were processed in two parallel pipelines. For depth‑image analysis, each raster was converted to an 8‑bit grayscale height‑map (ground‑referenced, normalized to 0–255) and padded to a fixed resolution. For point‑cloud analysis, the same rasters were back‑projected into 3‑D space using calibrated intrinsic parameters and a known camera‑to‑ground distance, discarding points below the ground plane. This unified conversion ensured that both representations originated from identical raw measurements.

Four experimental settings were defined to assess the impact of data representation: (1) unsegmented raw data, (2) segmented full‑body data, (3) segmented hindquarter (the region most directly linked to BCS), and (4) handcrafted geometric features derived from anatomical landmarks. Segmentation was achieved by first applying Grounded SAM with a “cow” text prompt to a subset of 130 images, manually refining the masks, and then training a lightweight YOLOv11n‑seg model. The trained model generated masks for all 1,020 cows. Keypoint detection (six primary landmarks: left/right short rib, left/right hook, left/right pin) was performed with a YOLOv11s‑pose network trained on 877 cows; performance was quantified by Percentage of Correct Keypoints (PCK) and Root‑Mean‑Square Error (RMSE).

Handcrafted features were extracted from the nine refined keypoints (including three derived mid‑points) by constructing ten lines (L1–L10) and computing, for each line, maximum protrusion distance, surface area, and volumetric measures (four sub‑volumes V1–V4). For depth images these calculations used 2‑D geometry; for point clouds the same lines and planes were mapped into true 3‑D space, preserving geometric fidelity.

Three families of predictive models were evaluated under each setting: (a) deep‑learning convolutional networks for depth images (ResNet‑18 and ConvNeXt), (b) deep‑learning point‑cloud networks (PointNet and DGCNN), and (c) traditional machine‑learning regressors for handcrafted features (Random Forest and LightGBM). All models were trained, validated, and tested within a cow‑level cross‑validation framework to avoid data leakage, using identical train‑validation‑test splits across modalities.

Results showed a clear advantage for depth‑image models when using unsegmented raw data and segmented full‑body data: both ResNet‑18 and ConvNeXt achieved lower RMSE (≈ 0.28) and higher R² than their point‑cloud counterparts. Point‑cloud models were more sensitive to sensor noise and to the choice of architecture; DGCNN exhibited larger performance variance than PointNet. When only the hindquarter was segmented, depth‑image and point‑cloud models performed comparably, suggesting that the hindquarter alone contains sufficient geometric cues for BCS estimation regardless of representation. Handcrafted feature models performed poorly for both data types, highlighting the loss of discriminative information when complex morphology is reduced to a few scalar descriptors.

Overall, the study concludes that, under the evaluated farm conditions, 3D point clouds do not provide a consistent performance boost over depth images for BCS prediction. Depth images remain a cost‑effective, robust, and high‑accuracy solution for automated BCS monitoring. The authors note that point‑cloud approaches may still hold promise if sensor noise can be reduced, segmentation accuracy improved, or hybrid architectures that fuse image and point‑cloud information are developed. Future work could explore higher‑resolution LiDAR, multi‑view fusion, or graph‑based networks tailored to livestock morphology to fully exploit the theoretical advantages of 3D point clouds.