An Improved Method for Broiler Weight Estimation Integrating Multi-Feature with Gradient Boosting Decision Tree

Body weight is an operationally important measure of broiler growth, flock uniformity, feeding decisions, and market readiness. Manual catching and weighing can be laborious, disruptive, and vulnerable to sampling bias. Camera-based estimation seeks to infer weight from visible morphology without requiring every measured bird to stand on a scale. Earlier approaches often used projected area, perimeter, major-axis length, or small sets of hand-engineered descriptors, with performance generally becoming more difficult as birds age and scenes become crowded.

Li et al. address two related problems: isolating individual birds in depth images containing complex backgrounds and representing enough body information to predict differences among same-age birds. Their premise is that engineered geometry remains useful, but that a convolutional network may capture additional patterns that a fixed feature list misses. Gradient-boosted trees are then used to combine these heterogeneous signals without assuming a simple linear relationship between morphology and weight.

The authors collected depth-distance and pseudo-color images with an Intel RealSense D435i. For the principal single-bird dataset, the camera was fixed above an electronic scale and five or six images were collected for each of 200 birds while ground-truth weight was recorded. A smaller multi-bird dataset was collected by holding the camera above groups of two or three chickens, followed by individual manual weighing. Mask R-CNN generated an instance mask so subsequent features could be calculated for each visible bird rather than for the whole frame.

The engineered feature set contains projected area, contour and convex-hull measures, major and minor axes, eccentricity, depth statistics, and surface-related descriptors. A customized ResNet50 produces a 2,048-dimensional learned representation. The authors compare engineered features, learned features, and their fusion using neural-network, XGBoost, and LightGBM regressors. Images were randomly shuffled into 70 percent training and 30 percent testing sets. The paper reports image-level metrics rather than an external barn, flock, or bird-level validation cohort.

Mask R-CNN segmentation was strong on the authors' held-out images. AP at an intersection-over-union threshold of 0.5 was 1.000 for single-bird and multi-bird training sets and 0.996 for the mixed set; average recall ranged from 0.872 to 0.900. These results show that the selected depth imagery allowed clean separation in this dataset. They should not be read as proof of equivalent segmentation under commercial occlusion, dust, changing litter, or different camera geometry.

For weight regression, engineered features alone with gradient boosting produced MAE near 0.13 kg and R-squared near 0.56. Learned features with the paper's larger neural network reached MAE of 0.104 kg and R-squared of 0.729. The lowest reported MAE, 0.093 kg, came from artificial and learned feature fusion with LightGBM, with R-squared of 0.707. The small difference between learned-only and fused configurations indicates that fusion improved absolute error in this experiment, but no single configuration dominated every metric.

The study makes a useful contribution by testing a complete depth-based workflow and by reporting ablations across feature families and regressors. Separating segmentation from weight estimation is methodologically clear, and the open code and dataset improve reproducibility. Including older birds addresses a real gap in a literature often concentrated on younger, less crowded broilers. The comparison also demonstrates that tree-based models can remain competitive when engineered and learned numerical features are combined.

The evaluation unit requires caution. Each bird contributed several images, yet the paper describes a random image shuffle rather than an explicitly bird-disjoint split. If images of the same bird appear in training and testing, visual similarity may make the reported test error optimistic. The collection also places birds above a scale during principal imaging, producing cleaner geometry and ground truth than a fully passive barn installation. These design choices are appropriate for a baseline study but weaken claims about unattended daily flock estimation.

The data represent one breed type, one age, one facility, and a narrow collection period. The authors identify posture as a source of abnormal estimates and propose future posture detection and multi-camera data. They also note that manual collection is still required during model training. Cross-age performance is not demonstrated, despite growth stage being central to commercial use. A model trained at 63 days may not preserve calibration at earlier ages when body proportions, feathering, density, and occlusion differ.

External validation is absent. The reported MAE cannot be assumed for a different camera, mounting height, breed, litter surface, lighting condition, barn, or weight distribution. Uniformity estimation introduces another layer: even acceptable individual error can bias flock percentiles if the system preferentially sees active, isolated, or centrally located birds. A commercial evaluation should therefore report bird-disjoint and barn-disjoint error, bias by weight class, coverage, missingness, repeatability, and population-level uniformity error.

The paper supports a staged research architecture for nesteye: calibrate camera geometry, isolate reliable bird instances, compute physical and learned features, reject unsuitable poses, and estimate weight only when measurement quality is sufficient. It also supports treating daily weight visibility as a distributional problem rather than displaying one unexplained number. Coverage, confidence, and the number of valid observations should accompany any flock estimate.

The reported 0.093 kg MAE is Li et al.'s result on their dataset, not a nesteye result. The paper uses a depth camera and collection protocol that may differ from nesteye hardware. It does not validate daily accuracy, flock uniformity, or performance in a commercial broiler barn. Nesteye would need independent scale-linked ground truth, bird- and barn-disjoint validation, drift checks across the growth cycle, and explicit reporting of when the model declines to estimate.

Li et al. demonstrate that depth-based instance segmentation and multi-feature regression can estimate the weight of older bantam chickens with useful accuracy in a constrained dataset. The work is most persuasive as a method study: it identifies a concrete feature pipeline, compares alternative regressors, and shows where fusion changes error. It is less persuasive as evidence of general commercial readiness because the validation remains within a narrow image collection.

For research planning, the paper clarifies both opportunity and burden. Non-contact weight estimation is technically plausible, but trustworthy daily visibility depends on calibration, pose quality, representative sampling, and independent validation across birds, ages, and barns. Those conditions should define the evidence standard for any product claim derived from this line of work.