Sub-Saharan Africa Wildlife Detector: YOLO Weights for 36-Class Camera Trap Object Detection

Here's the complete, final Zenodo record: Title: Sub-Saharan Africa Wildlife Detector: YOLO Weights for 36-Class Camera Trap Object Detection Description: An open-source object detection model for Sub-Saharan African wildlife in camera trap and field imagery. A YOLO-family detector (YOLOv10x / YOLO26x variants, performance comparable between the two) trained to localise 36 classes in a single pass: African mammal and bird species relevant to conservation monitoring, plus Person and Car utility classes. Weights are released in ONNX format for offline, cross-platform inference. This is a companion release to the UK Wildlife Detector, applying the same open, non-commercial approach to a Sub-Saharan African species set. Classes (36): Cheetah (Acinonyx jubatus jubatus), Car, Blue wildebeest (Connochaetes taurinus), Plains zebra (Equus quagga), Baboon (Papio sp.), Giraffe (Giraffa camelopardalis), African elephant (Loxodonta africana), Lion (Panthera leo), Person, Rhino, African buffalo (Syncerus caffer), Eland (Tragelaphus oryx), Spotted hyena (Crocuta crocuta), Impala (Aepyceros melampus), Black-backed jackal (Canis mesomelas), Sable antelope (Hippotragus equinus), Rabbit (Oryctolagus cuniculus), Warthog (Phacochoerus africanus), Chimpanzee (Pan troglodytes), Giant pangolin (Smutsia gigantea), Crested porcupine (Hystrix cristata), Aardvark (Orycteropus afer), Hippopotamus (Hippopotamus amphibius), Gemsbok (Oryx gazella), Ostrich (Struthio camelus), Hartebeest (Alcelaphus buselaphus), Waterbuck (Kobus ellipsiprymnus), Gorilla (Gorilla sp.), Bongo (Tragelaphus eurycerus), Kob (Kobus kob), Helmeted guineafowl (Numida meleagris), Common duiker (Sylvicapra grimmia), Genet (Genetta sp.), Giant pouched rat (Cricetomys sp.), Vervet monkey (Chlorocebus tantalus), Leopard (Panthera pardus). Training data: The dataset comprises 57,350 labelled instances split 80/10/10 into training (45,879), validation (5,735), and test (5,736) sets. Imagery spans daylight and infrared night-time camera trap frames, alongside safari-park/zoo field deployments, across a range of habitats, lighting conditions, and animal poses. Training setup: Architecture: YOLO-family extra-large variant (YOLOv10x and YOLO26x releases available); Hardware: 8× GPUs, global batch size 256; Input resolution: 640×640, mixed-precision (AMP); Epochs: 65; Optimiser: initial LR 0.01, final LR factor 0.01, momentum 0.937, weight decay 0.0005, 3-epoch warmup; Augmentation: mosaic (closed last 10 epochs), horizontal flip p=0.5, HSV jitter (h=0.015, s=0.7, v=0.4), random affine (scale ±0.5, translate ±0.1), RandAugment, random erasing 0.4; Loss weights: box=7.5, cls=0.5, dfl=1.5; Framework: Ultralytics/PyTorch, seed=0, deterministic=true; Export format: ONNX Performance (validation split): Metric Value mAP@0.5 (all classes) 0.980 Precision (at confidence 1.0) 1.00 Recall (at confidence 0.0) 0.99 F1 (peak, all classes) 0.96 at confidence 0.467 Training converged cleanly over 65 epochs with no late-stage overfitting. The confusion matrix shows a clean diagonal with minimal cross-class confusion; residual error is concentrated in background false positives rather than species-to-species confusion. Intended use: A single-stage, species-level object detector for African conservation practitioners, ecologists, and citizen science projects. Free for non-commercial conservation, research, teaching, and citizen science use. Related publication: No accompanying paper yet. This release is documented via this Zenodo record and the included training artefacts (confusion matrices, PR/F1/precision/recall curves, training logs, label distribution). License: CC BY-NC 4.0 Citation: Please cite this Zenodo DOI when using this model in published work. Commercial use requires separate permission.