Skip to content

Model Analysis Report

Hub slug rotated_rtmdet_dota_le90_3x. Official DOTA v1.0 Task 1 77.60% (AP75 52.01, COCO mAP 49.02). Deploy production.score_threshold 0.45 (this sweep’s best F1 0.50 − 0.05). The mAP50 below is leaky eval-val (val tiles are in train), not the published number. eval-val − Task 1 is 5.15 (82.75 − 77.60). Metrics margin 0 (all 57,768 val objects).

Official Task 1 (hidden test)

Class AP50
plane 0.8923
baseball-diamond 0.8189
bridge 0.5348
ground-track-field 0.7288
small-vehicle 0.8104
large-vehicle 0.8290
ship 0.8885
tennis-court 0.9069
basketball-court 0.8623
storage-tank 0.8660
soccer-ball-field 0.6101
roundabout 0.6783
harbor 0.7725
swimming-pool 0.8256
helicopter 0.6159
mAP50 0.7760
  • AP75: 0.5201
  • COCO-style mAP: 0.4902
  • Description submitted: oriented-det pretrained weights for RTMDet 3x (correct alias)
  • Experiment: runs/rotated_rtmdet/20261002-102116 (best_mAP_0.88.pth)

  • Recipe: configs/rotated_rtmdet/dota_le90_3x_s.json

  • Generated at: 2026-10-03T04:35:35.234909

Model metadata

  • Experiment dir: runs/rotated_rtmdet/20261002-102116
  • Checkpoint: runs/rotated_rtmdet/20261002-102116/checkpoints/best_mAP_0.88.pth
  • Checkpoint modified: 2026-10-02T18:26:26.554656
  • Config: runs/rotated_rtmdet/20261002-102116/config.json

Source data

  • Data root: /path/to/data/DOTA-v1.0-tiled
  • Data split: val
  • Total images: 7669
  • Total ground truth objects: 57768
  • Total predictions: 247861

Evaluation setup

  • mAP / PR matching IoU (rotated boxes, VOC-style; not NMS IoU): 0.50
  • NMS IoU (deduplication): 0.10
  • Threshold sweep: 0.0 to 1.0 step 0.05

Key outcomes

  • Best threshold (F1): 0.5000
  • Precision at best threshold: 0.8114
  • Recall at best threshold: 0.9236
  • F1 at best threshold: 0.8639
  • F2 at best threshold: 0.8987
  • mAP50: 0.8275 (82.75%)

GT alignment (mean best IoU vs raw detections)

  • Global mean best IoU (any class): 0.8022
  • Global mean best IoU (same class): 0.8007 (median 0.8277)

Per-class breakdown (each GT: max rotated IoU vs detections on the same image):

Class gts mean_any mean_same med_same
baseball-diamond 364 0.7990 0.7988 0.8111
basketball-court 278 0.8847 0.8843 0.8981
bridge 666 0.6818 0.6811 0.7229
ground-track-field 216 0.8322 0.8293 0.8634
harbor 4298 0.7619 0.7609 0.7996
helicopter 157 0.7540 0.7527 0.7872
large-vehicle 9398 0.8278 0.8225 0.8429
plane 4731 0.8392 0.8392 0.8731
roundabout 256 0.8057 0.8056 0.8414
ship 18534 0.8115 0.8109 0.8273
small-vehicle 11357 0.7758 0.7749 0.7981
soccer-ball-field 260 0.7868 0.7649 0.8757
storage-tank 5031 0.7747 0.7745 0.8338
swimming-pool 693 0.6817 0.6814 0.7179
tennis-court 1529 0.9122 0.9114 0.9237
global 57768 0.8022 0.8007 0.8277

Per-class metrics (mAP50)

Class gts dets recall AP
baseball-diamond 364 4270 0.981 0.7755
basketball-court 278 2285 1.000 0.9658
bridge 666 19886 0.884 0.6717
ground-track-field 216 3437 0.977 0.7891
harbor 4298 16963 0.950 0.8508
helicopter 157 1096 0.949 0.8998
large-vehicle 9398 48136 0.983 0.8892
plane 4731 12983 0.969 0.8981
roundabout 256 3558 0.965 0.7930
ship 18534 49208 0.986 0.7361
small-vehicle 11357 51593 0.977 0.8853
soccer-ball-field 260 1936 0.881 0.7802
storage-tank 5031 24317 0.917 0.8594
swimming-pool 693 4331 0.893 0.7488
tennis-court 1529 3807 0.995 0.8701
mAP 0.8275

Per-class best thresholds (max F1 over the same sweep)

Class Threshold Precision Recall F1 TP FP FN
baseball-diamond 0.6000 0.7260 0.8736 0.7930 318 120 46
basketball-court 0.7000 0.9306 0.9640 0.9470 268 20 10
bridge 0.4500 0.6241 0.7402 0.6772 493 297 173
ground-track-field 0.6000 0.7720 0.8935 0.8283 193 57 23
harbor 0.4500 0.8474 0.9118 0.8784 3919 706 379
helicopter 0.4500 0.9221 0.9045 0.9132 142 12 15
large-vehicle 0.6000 0.9154 0.9265 0.9209 8707 805 691
plane 0.6000 0.9406 0.9379 0.9392 4437 280 294
roundabout 0.5500 0.7348 0.8984 0.8084 230 83 26
ship 0.5500 0.7136 0.9446 0.8130 17507 7026 1027
small-vehicle 0.5000 0.8913 0.9088 0.8999 10321 1259 1036
soccer-ball-field 0.6500 0.9032 0.7538 0.8218 196 21 64
storage-tank 0.5000 0.9104 0.8299 0.8683 4175 411 856
swimming-pool 0.5000 0.7895 0.7576 0.7732 525 140 168
tennis-court 0.5500 0.9221 0.9836 0.9519 1504 127 25

Confusion matrix

Computed at score threshold 0.5000 and IoU 0.50.

Rows are ground-truth classes; columns are predicted classes. The False Positive row contains unmatched detections; the Missed column contains unmatched GTs.

Actual \ Predicted baseball-diamond basketball-court bridge ground-track-field harbor helicopter large-vehicle plane roundabout ship small-vehicle soccer-ball-field storage-tank swimming-pool tennis-court Missed
baseball-diamond 339 0 0 0 0 0 0 0 0 0 0 0 0 0 0 25
basketball-court 0 274 0 0 0 0 0 0 0 0 0 0 0 0 4 0
bridge 0 0 429 0 0 0 0 0 0 0 0 0 0 0 0 237
ground-track-field 0 0 0 204 0 0 0 0 0 0 0 0 0 0 0 12
harbor 0 0 0 0 3763 0 0 0 0 5 0 0 0 0 0 530
helicopter 0 0 0 0 0 137 0 1 0 0 0 0 0 0 0 19
large-vehicle 0 0 0 0 0 0 8926 0 0 0 44 0 0 0 0 428
plane 0 0 0 0 0 0 0 4505 0 0 0 0 0 0 0 226
roundabout 0 0 0 0 0 0 0 0 233 0 0 0 0 0 0 23
ship 0 0 1 0 2 0 2 0 0 17799 2 0 0 0 0 728
small-vehicle 0 0 0 0 0 0 14 0 0 0 10320 0 0 0 0 1023
soccer-ball-field 0 0 0 0 0 0 0 0 0 0 0 205 0 0 0 55
storage-tank 0 0 0 0 0 0 0 0 1 0 0 0 4175 0 0 855
swimming-pool 0 0 0 0 0 0 0 0 0 0 0 0 0 525 0 168
tennis-court 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1512 16
False Positive 184 60 171 87 537 10 1322 391 99 7523 1214 43 411 140 142 0

Artifacts

  • Predictions JSON: predictions.json
  • Analysis JSON: analysis_iou0.50.json
  • PR curve: pr_curve.png
  • Threshold metrics: threshold_metrics.png

Notes

  • Global threshold selected by maximizing F1; tie-breaks favor recall, then lower threshold.
  • Per-class table: best threshold per class maximizes F1 on the same threshold grid (see best_threshold_per_class in the analysis JSON).
  • Precision/recall are computed using class-aware IoU matching with one-to-one assignment.