Comparative analysis of monocular depth generation and camera depth for non-invasive rabbit live weight estimation

• AI-generated depth maps outperform real depth and rival RGB for liveweight prediction. • Morphological pre-processing consistently degrades model performance across all input types. • MobileNetV4 achieves high accuracy with low computational cost for edge deployment. • Feature fusion (RGB + depth) shows no consistent advantage over unimodal models. • Generated depth enables sensor-free, scalable monitoring in precision rabbit farming. In contemporary commercial rabbit farming, accurate live weight assessment is essential for efficient production management. Traditional manual weighing is labour-intensive, stressful for animals, and impractical for frequent monitoring. This study evaluates computer vision approaches for non-invasive liveweight prediction by comparing RGB, real depth, and AI-generated depth images as input modalities. A dataset of synchronized RGB and depth images was collected from 1113 weaned and near-slaughter rabbits (1.2–7.7 kg) under controlled conditions. Image segmentation was performed using YOLOv11 without fine-tuning, demonstrating its feasibility for rabbit detection. Morphological pre-processing was applied to assess its impact, while tabular features from RGB were used to train traditional models (Extra Trees, Gradient Boosting). Deep learning models EfficientNetV2, MobileNetV4, and a dual-stream fusion model were trained on RGB, camera depth, and generated depth inputs. Results show that generated depth maps match or exceed the performance of RGB, with EfficientNetV2 achieving the highest accuracy using MG_4 (Marigold monocular depth model with 4 denoise steps)-an AI-generated depth map synthesized from RGB images (R² = 0.945, RMSE = 368.36 g). In our experiments, morphological pre-processing degraded performance across all tested models and input types, suggesting that such operations may remove features information that are informative for model prediction. The fusion model achieved competitive results (R² = 0.941, RMSE = 379.935 g), but gains were marginal compared to unimodal inputs. Notably, Extra Trees achieved R² = 0.945, RMSE = 366.241 g without pre-processing, rivalling deep learning models. MobileNetV4 offered the best efficiency-accuracy trade-off (31.18 MB, 1.79 GFLOPS), making it suitable for edge deployment. These findings highlight AI-generated depth as a scalable, cost-effective alternative to hardware depth sensors and advocate for pre-processing-free pipelines in precision rabbit farming.