Official code for "Self-Calibrated Shadow Detection with Spatial Consistency Constraints under Noisy Labels".
A training-time, backbone-agnostic framework that makes shadow detectors robust to label noise — without changing the raw labels and with zero extra inference cost. Built on SDDNet, it adds three training-only components:
- ICA — Iterative Confidence Aggregation: per-image EMA prediction history → down-weights temporally unreliable pixels.
- SCV — Spatial Consistency Verification: local boundary-structure agreement → down-weights spatially inconsistent regions.
- BRLF — Boundary Refined Loss Function: warm-up + per-epoch alternation between the original and reliability-weighted loss.
| Method | Venue | ISTD | SBU | UCF |
|---|---|---|---|---|
| FDRNet | ICCV'21 | 1.55 | 3.04 | 7.28 |
| SILT | ICCV'23 | 1.16 | 4.19† | 7.23† |
| SDDNet (baseline) | MM'23 | 1.27 | 2.94 | 6.59 |
| AdapterShadow | ESWA'25 | 0.86 | 2.75 | 6.35 |
| Ours | — | 1.14 | 2.70 | 6.20 |
Bold: best, underline: second best. †SILT evaluates SBU/UCF on its re-annotated SBU-Refine test set and is not directly comparable on those two columns.
Compared with the SDDNet baseline, our method reduces BER by 8.16% on SBU and 10.24% on ISTD, and reaches state-of-the-art BER on SBU and UCF.
@misc{xie2026sclshadow,
title = {Self-Calibrated Shadow Detection with Spatial Consistency Constraints under Noisy Labels},
author = {Xie, Jiaxuan and Chen, Xiao-Diao and Mo, Yuchang and Wu, Wen},
year = {2026}
}Released under the MIT License.
