SAR Ship Detection: RetinaNet vs YOLO11s vs RF-DETR Benchmark

Benchmarked three object detection architectures — a from-scratch RetinaNet, a fine-tuned YOLO11s, and a fine-tuned RF-DETR-Small — on the HRSID SAR ship detection dataset (1,962 test images: 369 inshore / 1,593 offshore), scoring all three with the exact same hand-written mAP implementation, validated against pycocotools, so the ranking reflects the models rather than three different eyeballed metrics.

RF-DETR-Small won on every split, reaching mAP@0.5:0.95 of 0.644 on the full test set versus 0.592 for YOLO11s and 0.485 for RetinaNet, despite training for only 20 epochs at a lower 512² resolution and with fewer parameters than RetinaNet. Its NMS-free, bipartite-matching design also gave it by far the strongest inshore recall (0.672 vs. 0.411 for YOLO11s and 0.304 for RetinaNet) in the dense, occluded harbor scenes that punish every model.

To sanity-check generalization beyond curated 800×800 chips, all three were also run on a full-size, 6720×9323px Sentinel-1 SAR scene of the Helsinki coast — a different sensor and location than anything in training, processed from raw GRD data through orbit correction, thermal/border noise removal, calibration, speckle filtering, and terrain correction in ESA SNAP before inference.

Full writeup, architecture diagrams, and honest caveats — mismatched training budgets, different input resolutions, single-seed results — are in the linked repo, alongside the from-scratch RetinaNet build and both fine-tuning baselines.