Overhead MegaDetector - OWL (Overhead Wildlife Locator) Benchmark - Models and Caribou Data

Point-annotated 512×512 px aerial image patches for caribou detection and counting from overhead survey imagery. This dataset accompanies the OWL paper and  enables reproducible evaluation of point-based object detectors on aerial wildlife imagery.

 Train split (PCH 2017): 23,517 patches (18,322 annotated with 273,268 point annotations + 5,195 background controls) from the Porcupine Caribou Herd, Alaska.

 Test split (CAH 2022): 2,607 patches (1,852 annotated with 12,456 point annotations + 755 background controls) from the Central Arctic Herd, Alaska.

 This is a strict cross-herd and cross-temporal generalization benchmark: models trained on PCH 2017 are evaluated on CAH 2022 without any per-deployment  retraining.

 Also includes the pre-trained caribou HerdNet (DLA-34) weights that reproduce the paper headline (F1 = 0.965 at τ = 20 px, c* = 0.20 on the test split),  together with the three OWL benchmark model checkpoints (OWL-C, OWL-T, OWL-D).

 Contents:

• test.zip — 2,607 test patches (512×512 PNG) + gt.csv (12,456 annotations)
• train.zip — 23,517 training patches (512×512 PNG) + gt.csv (273,268 annotations)
• Caribou-OWL-C.pth — Pre-trained caribou HerdNet best_model (DLA-34, epoch 14, val F1 = 0.937); previously distributed inside weights.zip
• OWL-C.pth — OWL-C benchmark model (HerdNet detection branch, DLA-34)
• OWL-T.pth — OWL-T benchmark model (HerdNet hybrid multi-scale residual)
• OWL-D.pth — OWL-D benchmark model (HerdNetDINO, frozen DINOv3 ViT-H+/16 backbone + DPT decoder)
• README.md — Dataset documentation, annotation format, benchmark results