Comparative evaluation of object-detection models for MMA fighter localisation in mixed martial arts bouts

Automated performance analytics in Mixed Martial Arts (MMA) depend on an upstream person-detection stage that feeds tracking, pose estimation, action recognition, strike counting, and the automated flagging of illegal manoeuvres. The in-cage environment of a UFC-style bout violates several assumptions of generic detectors trained on large pre-training corpora: high-velocity striking, body-on-body occlusion during clinch and grappling, cage-fence interference, dense crowd backgrounds, and a referee who is themselves a valid person detection to any COCO-pretrained model. This paper presents a preliminary, pilot-scale comparative evaluation of four state-of-the-art detectors - RF-DETR-XL, RF-DETR-XXL, DEIM D-FINE-X, and Ultralytics YOLO26x - on 250 broadcast-style frames at 1920×1080 and 640×360, with a single shared confidence threshold of 0.25. Predictions at the downscaled resolution are rescaled to native coordinates before evaluation. Because the ground truth labels only the two fighters per frame, each prediction XML is manually curated to remove non-fighter boxes; multi-box-per-fighter detections are retained so that this failure mode appears directly in the scored metrics. Localisation quality is reported as IoU-matched precision, recall, and F1 at IoU 0.5 and 0.75, complemented by torchmetrics COCO-style mAP (which is degenerate because Pascal VOC XML does not retain confidence scores; P/R/F1 is therefore the primary metric). Latency is measured as GPU-synchronised per-image time on a single NVIDIA RTX 5060 Laptop GPU, with the first frame excluded as warm-up. RF-DETR-XL at 640×360 delivers the strongest localisation ([email protected] = 0.933, [email protected] = 0.814) at ~13.5 FPS, while YOLO26x at 640×360 is the only configuration capable of real-time throughput at ~25.5 FPS, at a modest accuracy cost ([email protected] = 0.921, [email protected] = 0.776). All four models share a common failure mode in clinch and ground-grappling sequences, producing either multiple partial boxes per fighter or a single fused box covering both, motivating MMA-specific fine-tuning and a downstream tracker or region-of-interest stage for deployment.