Benchmarking Data and Trained YOLOv8-YOLO26 Machine Learning Models for SAR Ship Detection

All sizes of each generation of YOLO models, from YOLOv8n to YOLO26x were trained on the training split of the SAR Ship Detection Dataset (SSDD), and separately on the training split of the High Resolution SAR Images Dataset (HRSID). YOLOv8 YOLOv9 YOLOv10 YOLO11 YOLO12 YOLO26 YOLOv8n YOLOv9t YOLOv10n YOLO11n YOLO12n YOLO26n YOLOv8s YOLOv9s YOLOv10s YOLO11s YOLO12s YOLO26s YOLOv8m YOLOv9m YOLOv10m YOLO11m YOLO12m YOLO26m YOLOv8l YOLOv9c YOLOv10b YOLO11l YOLO12l YOLO26l YOLOv8x YOLOv9e YOLOv10l YOLO11x YOLO12x YOLO26x YOLOv10x All models were trained on a Dell Precision workstation with Intel Xeon W-2255 CPU, 192GB DDR4 memory, and Dual Nvidia RTX A4000 GPUs with 16GB DDR4 VRAM using identical hyperparameters for consistency, using the AdamW optimiser for 500 epochs, with close_mosaic=350, scale=0.5, fliplr=0.5, flipud=0.5, cos_lr=True, lr0=0.001, lrf=0.01, momentum=0.937, weight_decay=0.0005, and degrees=0. Checkpoints were saved for every training epoch, totalling 31,000 checkpoints: 500 checkpoints for each of 31 models for both training datasets. All 31,000 checkpoints were tested on three separate test sets, to assess in-domain and cross-domain performance: 1. The test split of the SSDD 2. The test split of the HRSID 3. A test split of ther Large Scale SAR Ship Detection Dataset (LS-SSDD) This produced 3 test results for each of the 500 epochs for each of the 31 models for each of the 2 training datasets; a total of 93,000 test results. This dataset comprises these 93,000 test results, along with .pt model files for representative epochs for each size/generation of YOLO. Test results are presented in two CSV files: one for models trained on the SSDD, and the second for models trained on the HRSID. Both are formatted identically. Columns: checkpoint str: the model, size, training dataset, and epoch, e.g. yolov8n_SSDD_epoch0 indicates YOLOv8 nano trained on SSDD, epoch 0family str: the generation of YOLO model, YOLOv8, YOLOv9, YOLOv10, YOLO11, YOLO12, YOLO26size str: the size of model, Nano, Tiny, Small, Medium, Balanced, Large, Compact, Extra Large, Extensive letter str: the letter indicating the model size, n, t, s, m, b, l, c, x, emodel str: combination of the generation and size e.g. YOLOv8nGFLOPs float: computational resources in GFLOPstrained_on str: SSDD or HRSIDepoch int: zero-indexed training epoch number (0-499)model_size str: the size category of model, combining variant names, Nano/Tiny, Small, Medium/Balanced, Large/Compact, Extra Large/Extensive family_sort int: the YOLO generation number to allow sorting, 8, 9, 10, 11, 12, 26size_sort int: size categories numbered 1-5 to allow sorting, 1= Nano/Tiny, 2 = Small, 3 = Medium/Balanced, 4 = Large/Compact, 5 = Extra Large/Extensive model_sort int: models numbered 1-31 sorted by generation and sizemAP50_SSDD float: mAP50 metric tested on the test split of the SSDDmAP50-95_SSDD float: mAP50-95 metric tested on the test split of the SSDDPrecision_SSDD float: Precision metric tested on the test split of the SSDDRecall_SSDD float: Recall metric tested on the test split of the SSDDF1_SSDD float: F1 metric tested on the test split of the SSDDmAP50_HRSID float: mAP50 metric tested on the test split of the HRSIDmAP50-95_HRSID float: mAP50-95 metric tested on the test split of the HRSIDPrecision_HRSID float: Precision metric tested on the test split of the HRSIDRecall_HRSID float: Recall metric tested on the test split of the HRSIDF1_HRSID float: F1 metric tested on the test split of the HRSIDmAP50_LS-SSDD float: mAP50 metric tested on the test split of the LS-SSDDmAP50-95_LS-SSDD float: mAP50-95 metric tested on the test split of the LS-SSDDPrecision_LS-SSDD float: Precision metric tested on the test split of the LS-SSDDRecall_LS-SSDD float: Recall metric tested on the test split of the LS-SSDDF1_LS-SSDD float: F1 metric tested on the test split of the LS-SSDD Trained models provided include the checkpoint with the highest in-domain mAP for each YOLO generation/size. Epoch with max in-domain mAP for SSDD Epoch with max in-domain mAP for HRSID yolov8n 340 495 yolov8s 264 242 yolov8m 232 219 yolov8l 267 166 yolov8x 190 206 yolov9t 256 195 yolov9s 325 232 yolov9m 214 112 yolov9c 150 201 yolov9e 239 184 yolov10n 394 335 yolov10s 434 210 yolov10m 448 447 yolov10b 234 143 yolov10l 227 499 yolov10x 443 277 yolo11n 365 453 yolo11s 406 418 yolo11m 240 290 yolo11l 219 165 yolo11x 238 226 yolo12n 298 334 yolo12s 312 327 yolo12m 391 227 yolo12l 388 290 yolo12x 319 287 yolo26n 284 268 yolo26s 312 235 yolo26m 445 223 yolo26l 350 206 yolo26x 402 316