Validation report
Completed on October 28, 2026 (Asia/Shanghai).
Paper, historical, and re-evaluated results
The score order is the same as in the model card. All scores are validation mIoU (%); ACDC is the mean over four conditions.
| Model ID | Paper | Historical log | Re-evaluated |
|---|---|---|---|
rdg_usg2cbm_dinov3_rein |
78.9 / 63.4 / 74.8 | 78.91 / 63.43 / 74.78 | 78.91 / 63.44 / 74.78 |
rdg_usg2cbm_dinov3-7b_rein |
80.0 / 66.3 / 75.9 | 80.02 / 66.32 / 75.91 | 80.02 / 66.32 / 75.91 |
rdg_c2abm_dinov3_rein |
72.5 / 65.0 / 77.4 | 72.52 / 65.04 / 77.42 | 72.52 / 65.04 / 77.41 |
rdg_c2abm_dinov3-7b_rein |
75.9 / 68.0 / 77.5 | 75.92 / 67.98 / 77.54 | 75.92 / 67.97 / 77.54 |
rdg_g2cbm_dinov3-7b_rein |
73.9 / 65.5 / 72.6 | 73.90 / 65.49 / 72.58 | 73.89 / 65.49 / 72.58 |
rda_g2c_dinov3_rein_ta |
79.0 | 79.03 | 79.03 |
Twelve of the sixteen reported metrics match the historical logs at two decimal places. Four differ by 0.01 percentage points. All match the paper at one decimal place. The cause of the small differences has not been isolated. The complete per-domain metrics, including the Cityscapes source-domain diagnostic scores for c2abm, are in evaluations.json.
Forward equivalence
All six models were compared against the original experiment implementation in FP32 with TF32 disabled. Each comparison used a real image at 768 × 512 and a fixed random input at 736 × 528. In all twelve comparisons, the maximum absolute logit difference was 0.0 and predicted labels were identical. These are paired-input checks, separate from the full validation-set evaluations above.
The per-model JSON reports in validation/ record input, source-code, configuration, and checkpoint hashes. All six full dataset evaluations completed successfully.
Additional checks
- Configuration and parameter-loading preflight: passed for all six models.
- Automated tests: 19 passed.
- Single-image inference CLI: passed for
rdg_usg2cbm_dinov3_rein; mask and overlay outputs checked. - Missing dataset path: expected error confirmed.
- CUDA deformable attention against its PyTorch reference: maximum absolute error 1.1920928955078125e-07.
- Label conversion: 10 Cityscapes and 32 Mapillary masks matched the original evaluation masks pixel for pixel.
- Release gate: all six tested weight/configuration hashes match the staged files, and the Python source hash matches the code used in GPU validation.
The tested environment is Linux/Python 3.10, PyTorch 2.1.0+cu118, MMCV 2.1.0, MMEngine 0.10.7, MMSegmentation 1.2.2, and MMDetection 3.3.0. No official test-server evaluation was performed in this run.