The present study provides a systematic benchmarking of nine state-of-the-art object detection models adapted to a specialized animal dataset. The main objective was to evaluate their performance in terms of accuracy and inference speed in the context of real-time and commercial applications within the agro-industrial sector. The method-ology involved aggregating and thoroughly cleaning several cattle image datasets, fol-lowed by fine-tuning models with diverse architectures — two-stage, one-stage, and transformer-based. Performance evaluation was conducted on a unified hardware platform equipped with an NVIDIA RTX 4090 GPU with 24 GB of VRAM, which is critical for the objec-tive assessment of speed-related metrics. Key results show that the D-FINE and Co-DETR models demonstrated the highest ac-curacy (AP@[IoU=0.50:0.95] of 0.872 and 0.851, respectively), while RTMDet and YOLOv11 proved to be the fastest (15.81 ms/image and 19.14 ms/image in reproducible mode). A general trend was observed toward significant improvement in average pre-cision (AP) metrics on the specialized dataset compared to their known results on gen-eral-purpose datasets such as COCO, while relative inference speed rankings remained largely consistent.

This study benchmarks nine state-of-the-art object detection models on a specialized cattle dataset to assess accuracy and inference speed for real-time agricultural applications. Using a unified protocol without model-specific augmentations, and evaluating all detectors on identical RTX 4090 hardware, we provide a fair architectural comparison of two-stage, one-stage, and transformer-based models. D_FINE_L and Co_DETR_R_50 achieved the highest accuracy (AP@[0.50:0.95] = 0.872 and 0.851), while RTMDet and YOLOv11_L were the fastest (15.81 and 19.14 ms/image). All models showed substantial accuracy gains on the domain dataset compared to COCO, while maintaining consistent relative speed rankings.

This paper presents the design of a hardware–software system for non-invasive automated monitoring of feeding behavior in livestock with biometric identification of individual animals. Neural network models for animal identification from images and individual recognition have been developed and trained. A solution is proposed to address the challenge of acquiring a sufficient number of personalized animal images for training the identification neural network. A transfer learning approach is introduced for pig identification, where the network is first trained on a large-scale dataset of more than three million human face images obtained from open sources and subsequently fine-tuned by training the upper layers on a significantly smaller dataset consisting of 5610 pig face images. Experimental results demonstrated the high effectiveness of the system: the Top-1 identification accuracy reached 95.1%, while the ROC AUC in open-set recognition tasks achieved 0.95. The processing time per frame on an NVIDIA RTX 4090 GPU was 1.4 ms (724 FPS).