Steiger β Blackwell
RF-DETR Seg 2XLarge instance segmentation for one class, steiger, in RGB aerial imagery. Input: 1008 Γ 1008 pixels, imagery resolution: 8 cm/pixel.
First run, using the original dataset and a step learning-rate schedule.
Checkpoint: blackwell-001-checkpoint_best_ema.pth Β· Selected epoch: 9 Β· Run started: 2026-09-09 Β· W&B run
Performance comparison
| Model | Validation data | Mask mAP 50β95 β | Mask mAP 50 β | Box mAP 50β95 β | EMA epoch |
|---|---|---|---|---|---|
| Blackwell | Original | 44.10% | 71.29% | 57.14% | 9 |
| Background | With backgrounds | 42.83% | 69.95% | 55.35% | 4 |
| Cosine | With backgrounds | 43.54% | 71.06% | 56.75% | 8 |
Cosine is the strongest of the two runs evaluated on the same background dataset, improving mask mAP 50β95 by 0.71 percentage points over Background. Blackwell has the highest recorded score, but used a different dataset version; its score is not a controlled comparison with the other two.
Scores are validation results for the best eligible EMA epoch, not final-epoch results. Epochs are numbered from 1. Higher is better; mask mAP 50β95 is the primary metric. These are checkpoint-selection results, not an independent test-set benchmark.
The first three epochs were excluded by skip_best_epochs=3. The checkpoint fileβs stored epoch matches the selected W&B EMA epoch. Its SHA-256 checksum, exact scores, and source fields are recorded in metrics.json.
Validation examples
Four fixed validation examples from epoch 18. Green contours: ground truth; thick red contours: live-model predictions, with confidence β₯ 0.30 (up to 50 predictions per image). Red outlines have been thickened and given a white halo for visibility; the underlying predictions are unchanged. These previews are not EMA-checkpoint inference and are not a representative benchmark.
This gallery comes from an available logged epoch, which differs from the selected EMA epoch above.







