---
license: apache-2.0
---

# SFM: Spatial Frequency Modulation for Semantic Segmentation (TPAMI 2025)

Official checkpoints for the paper **Spatial Frequency Modulation for Semantic Segmentation**
(TPAMI 2025, [arXiv:2507.11893](https://arxiv.org/abs/2507.11893),
DOI [10.1109/TPAMI.2025.3592621](https://doi.org/10.1109/TPAMI.2025.3592621)).

Code: **https://github.com/Linwei-Chen/SFM**

Two released technical lines:

- **Line 1 — FCN-style models**: ARS + qhull non-uniform upsampling (MSAU, paper Eq. 17/18)
  + cascaded LPRM, built on mmsegmentation 0.25.0.
  Eval code and the 18 configs: [`SFM_mmseg0x/`](https://github.com/Linwei-Chen/SFM/tree/main/SFM_mmseg0x).
- **Line 2 — FPN**: ARS + CARAFE-style feature reassembly with 3-point dynamic-sampling
  weighted alignment, on Mask2Former (mmsegmentation 1.x).
  Code and config: [`SFM_mask2former/`](https://github.com/Linwei-Chen/SFM/tree/main/SFM_mask2former).

Every checkpoint below is **verified from training/eval logs**: each measured mIoU either
exceeds the paper-reported number or matches it after one-decimal rounding
(e.g. PSPNet-R50 D16 measures 78.05, rounding to the paper's 78.1). Each model folder
contains `config.py` (the exact config paired with that checkpoint) and the weight file.

## Line 1 — Cityscapes (val mIoU, ResNet-50)

| folder | paper | measured |
| --- | --- | --- |
| `line1/cityscapes/ccnet_r50SS3PR-d32_cityscapes` | 76.3 | **76.34** |
| `line1/cityscapes/ccnet_r50SS3PR-d16_cityscapes` | 78.5 | 78.45 |
| `line1/cityscapes/ccnet_r50SS3PR-d8_cityscapes` | 79.3 | **79.39** |
| `line1/cityscapes/isanet_r50SS3PR-d32_cityscapes` (OCNet) | 75.8 | **75.82** |
| `line1/cityscapes/isanet_r50SS3PR-d16_cityscapes` (OCNet) | 78.0 | **78.00** |
| `line1/cityscapes/isanet_r50SS3PR-d8_cityscapes` (OCNet) | 79.7 | **79.74** |
| `line1/cityscapes/FCN_r50SS3PR-d32_cityscapes` | 76.2 | **76.20** |
| `line1/cityscapes/FCN_r50SS3PR-d16_cityscapes` | 77.8 | 77.78 |
| `line1/cityscapes/FCN_r50SS3PR-d8_cityscapes` | 78.7 | 78.65 |
| `line1/cityscapes/pspnet_r50SS3PR-d32_cityscapes` | 76.0 | **76.18** |
| `line1/cityscapes/pspnet_r50SS3PR-d16_cityscapes` | 78.1 | 78.05 |
| `line1/cityscapes/pspnet_r50SS3PR-d8_cityscapes` | 79.5 | **79.52** |
| `line1/cityscapes/pcaa_r50SSPR-d32_cityscapes` (PCAA, col. 1) | 76.8 | **76.83** |
| `line1/cityscapes/pcaa_r50SSPR-d16_cityscapes` (PCAA, col. 2) | 77.9 | 77.86 |

Weight file in each folder: `iter_40000.pth`.

## Line 1 — PascalContext-59 (val mIoU, ResNet-50)

| folder | paper | measured |
| --- | --- | --- |
| `line1/pascalcontext/FCN_r50SS3PR-d32_pascalcontext` | 46.0 | **46.35** |
| `line1/pascalcontext/pspnet_r50SS3PR-d32_pascalcontext` | 49.7 | 49.68 |
| `line1/pascalcontext/ccnet_r50SS3PR-d32_pascalcontext` | 49.1 | 49.07 |
| `line1/pascalcontext/isanet_r50SS3PR-d32_pascalcontext` (OCNet) | 49.9 | 49.85 |

## Line 2 — ADE20K (val mIoU, Mask2Former Swin-T)

| folder | paper | measured |
| --- | --- | --- |
| `line2/mask2former_swin-t_8xb2-160k_ade20k-512x512_SFMv2` | 49.1 | **49.34** (re-evaluated 2025-10) |

Weight file: `best_mIoU_iter_150000.pth`.

## Evaluation

**Line 1** (see [`SFM_mmseg0x/README.md`](https://github.com/Linwei-Chen/SFM/blob/main/SFM_mmseg0x/README.md)
for environment setup and qhull compilation):

```bash
cd SFM_mmseg0x    # from the GitHub repo checkout
PYTHONPATH=<SFM_mmseg0x> python tools/test.py \
    <model_folder>/config.py <model_folder>/iter_40000.pth --eval mIoU
```

**Line 2** (code: [`SFM_mask2former/`](https://github.com/Linwei-Chen/SFM/tree/main/SFM_mask2former),
mmsegmentation 1.x):

```bash
cd SFM_mask2former
python tools/test.py \
    <model_folder>/config.py <model_folder>/best_mIoU_iter_150000.pth
```

## Citation

```bibtex
@article{chen2023spatialfrequency,
  title={Spatial Frequency Modulation for Semantic Segmentation},
  author={Chen, Linwei and Fu, Ying and Gu, Lin and Zheng, Dezhi and Dai, Jifeng},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2025},
  volume={47},
  number={11},
  pages={9767-9784},
  doi={10.1109/TPAMI.2025.3592621}
}
```

Code is released under the Apache-2.0 license.