AutoThreshold: Joint Confusion-Profile Learning for Adaptive Mask Binarization
Organizations: Shanghai Jiao Tong University · Zelos Technology
Abstract
A single validation-tuned threshold can become brittle when dense-score distributions shift. We introduce AutoThreshold, a post-hoc policy that predicts the joint mass of positive and negative pixels over score bins. This target directly parameterizes threshold-indexed confusion counts, avoiding products of independently estimated prevalence and class-conditional rates. Softmax normalization and cumulative decoding yield coherent, monotone confusion profiles; a plug-in decoder then selects a threshold for any objective without retraining the head. Across three independently trained Kvasir pipelines transferred to four polyp datasets, Joint exceeds continuous Otsu by external macro Dice, with positive gains in 3/3 runs. On the reference pipeline, three head seeds achieve , 0.0085 above Otsu; direct-utility Ridge achieves 0.6342, with the paired difference interval spanning zero. The same Joint checkpoints outperform Otsu for , Dice, and . On developmental KITTI Road leave-one-corruption-family-out evaluation, Joint improves over scalar regression by 0.0082 [0.0064, 0.0101], while direct utility is 0.0022 higher. These results establish joint confusion prediction as a useful, metric-retargetable inductive bias, without claiming Bayes-optimal decisions.
Figures & tables
| Dataset | N | Fixed .5 | Global | Otsu | Target global † | Joint | Oracle |
|---|---|---|---|---|---|---|---|
| Held-out source-domain test | |||||||
| Kvasir-SEG | 200 | 0.8903 | 0.8902 | 0.8912 | 0.8904 | 0.8885 | 0.9238 |
| External target datasets without target tuning | |||||||
| CVC-300 | 60 | 0.7873 | 0.7876 | 0.7994 | 0.7883 | 0.8106 | 0.8632 |
| CVC-ClinicDB | 62 | 0.6566 | 0.6537 | 0.6603 | 0.6655 | 0.6691 | 0.7698 |
| CVC-ColonDB | 380 | 0.5583 | 0.5584 | 0.5826 | 0.5595 | 0.5837 | 0.6764 |
| Method | Kvasir test | External |
|---|---|---|
| Fixed 0.5 | 0.8903 | 0.6087 |
| Global threshold | 0.8902 | 0.6080 |
| Otsu [ 21 ] | 0.8912 | 0.6239 |
| Area-only linear (scalar) | 0.8904 | 0.6087 |
| Ridge (scalar, 61 features) | 0.8879 | 0.6014 |
| GBRT (scalar, 61 features) | 0.8899 | 0.6073 |
| Method | Dice/ | ||
|---|---|---|---|
| Global | 0.6250 | 0.6080 | 0.6351 |
| Otsu [ 21 ] | 0.6305 | 0.6239 | 0.6418 |
| Joint confusion | 0.6577 | 0.6324 | 0.6485 |
| Oracle | 0.7328 | 0.7151 | 0.7483 |
| Split/model seed | Global | Otsu | Direct utility | Joint |
|---|---|---|---|---|
| 20260726 | 0.6080 | 0.6239 | 0.6279 | 0.6351 |
| 20260727 | 0.5941 | 0.6078 | 0.6130 | 0.6105 |
| 20260728 | 0.6286 | 0.6458 | 0.6531 | 0.6558 |
| Method | Dice | IoU |
|---|---|---|
| Clean images | ||
| Fixed 0.5 | 0.8795 | 0.7985 |
| Otsu [ 21 ] | 0.8790 | 0.7977 |
| Validation-tuned global | 0.9067 | 0.8376 |
| Scalar regression | 0.9072 | 0.8398 |
| Per-image oracle | 0.9142 | 0.8512 |
| Family | Scalar | Direct | Joint | Aware † | J S | J D |
|---|---|---|---|---|---|---|
| Brightness | 0.9029 | 0.9098 | 0.9076 | 0.9073 | +0.0048 | 0.0022 |
| Gamma | 0.8949 | 0.9005 | 0.8979 | 0.8969 | +0.0029 | 0.0026 |
| Gaussian blur | 0.8140 | 0.8532 | 0.8493 | 0.8492 | +0.0353 | 0.0039 |
| JPEG | 0.9016 | 0.9029 | 0.9030 | 0.8944 | +0.0014 | +0.0001 |
| Low contrast | 0.9013 | 0.9058 | 0.9034 | 0.9031 | +0.0021 | 0.0024 |
| White veil | 0.8991 | 0.9043 | 0.9021 | 0.9010 | +0.0030 | 0.0023 |
Image Segmentation
Image segmentation assigns pixels or voxels to objects and regions, supporting measurement, scene understanding, and decisions across image domains.
See all 418 papers