SWS-TER

Sparse Weakly Semi-supervised Tri-Evidence Recovery Network
for Oriented Ship Detection in PolSAR Imagery

Yucong He · Gui Gao · Tianwen Zhang · Dunyun He

Official PyTorch implementation of the SWS-TER manuscript.

---

🚀 Overview

SWS-TER targets sparse, mixed-form supervision for oriented ship detection in PolSAR imagery. It is built on MMDetection and MMRotate and implements the complete three-stage training path described in the manuscript.

Overall architecture of the SWS-TER framework.

Main components

• ACPC (Stage I): polarimetric superpixels, revised Wishart saliency, region/background separability, four region anchors, diversity sampling, momentum SCFE, InfoNCE queue, prototype similarities and continuous P_tar/P_bg/P_hard priors
• SWCS (Stages II-III): scale-adaptive context compensation (SACC), prior-guided scattering keypoint graph (PSKG), GraphSAGE reasoning and evidence-guided context-scattering fusion (EGCSF)
• SALRP: ACPC-modulated sparse focal loss using Eq. (32), assigned by exact superpixel ownership and applied only to high-confidence negative FCOS locations in Eq. (33)
• UGSRT: per-FPN-level two-component GMM, teacher-student consistency, high/uncertain/low candidate partitioning, latent-feature mask tokens, online class prototypes and a lightweight ViT reconstructor
• Mixed weak annotations: RBox, HBox and Point supervision with center, superpixel/Voronoi, overlap and SAR edge-alignment constraints

The custom modules live under projects/SWS_TER, while the bundled legacy mmdet, mmrotate and semi_mmrotate packages provide the end-to-end training path.

---

🛠️ Installation

The tested compatibility path is Python 3.9, PyTorch 1.13.1, CUDA 11.6 and mmcv-full 1.7.1:

git clone https://github.com/YucongHe462/SWS-TER.git
cd SWS-TER
conda env create -f requirements/environment-legacy.yml
conda activate sws-ter-legacy

requirements/paper.txt records the environment used for the paper. See docs/REPRODUCIBILITY.md for compatibility notes and fixed experiment settings.

---

📦 Dataset

The dataset used in this work is publicly shared for benchmarking and reproducibility.

Shared download link: Download the SWS-TER dataset

📌 Access note: The hyperlink above is the complete shared-file URL. Dataset files are not stored in this Git repository.

---

⚙️ Data preparation and Stage I priors

Set SWS_TER_DATA_ROOT to the directory containing semi_ratio_20, test_image and test_annotation, then verify the released split:

$env:SWS_TER_DATA_ROOT = 'D:\datasets\Pd_Pv_Sa'
python tools\verify_Pd_Pv_Sa_data.py `
  --data-root $env:SWS_TER_DATA_ROOT `
  --expect-counts

To rebuild the split from Pascal VOC XML files instead:

python tools/prepare_gr_dataset.py \
  --images data/raw/sar_gray \
  --annotations data/raw/Annotations \
  --out-dir data/Pd_Pv_Sa \
  --seed 42

Run the annotation-free ACPC stage on all training images:

$TrainSplit = Join-Path $env:SWS_TER_DATA_ROOT `
  'semi_ratio_20\sparse_ratio_20'

python tools\acpc\run_acpc.py `
  --xpol-dirs (Join-Path $TrainSplit 'label_image') `
               (Join-Path $TrainSplit 'unlabel_image') `
  --work-dir work_dirs\acpc `
  --epochs 20 `
  --device cuda

The final work_dirs/acpc/acpc_priors.json file is consumed by the student head and PSKG support map. Optional polarimetric preparation utilities are retained for ablation studies but are not required by the single-channel path.

---

🚦 Training and evaluation

The released Pd_Pv_Sa 20%-image / 20%-instance HBox configuration is:

python tools/train.py configs/sws_ter/sws_ter_Pd_Pv_Sa_hbox_20_20.py \
  --work-dir work_dirs/sws_ter_Pd_Pv_Sa_hbox_20_20

python tools/test.py configs/sws_ter/sws_ter_Pd_Pv_Sa_hbox_20_20.py \
  work_dirs/sws_ter_Pd_Pv_Sa_hbox_20_20/latest.pth \
  --eval mAP

The published schedule is encoded in the config: 48,000 iterations, batch size 4, 12,800 burn-in iterations, 800 x 800 inputs, Adam at 5e-5, weight decay 0.05, linear warm-up and gradient clipping at norm 35. Inference uses the student branch.

For separate Stage II/Stage III commands and a short end-to-end runtime check, see docs/RUN_Pd_Pv_Sa.md.

---

✅ Verification

Core components have CPU-only unit tests and do not require MMCV CUDA ops:

pytest -q tests/test_core_modules.py
python -m compileall projects tools configs

The framework-level check requires the compiled legacy mmcv-full, but does not require dataset files or a checkpoint:

python tests/framework_smoke.py

It builds the full model and checks the single-channel forward path, UGSRT, Eq. (32) superpixel mapping, Eq. (33) gating, center-loss weight, student inference and model parameter counts.

---

📚 Documentation

• docs/METHOD_TO_CODE.md — equation-by-equation implementation map
• docs/REPRODUCIBILITY.md — settings, seeds and implementation choices
• docs/DATA.md — directory layout and annotation encoding
• docs/RUN_Pd_Pv_Sa.md — complete PowerShell commands for every stage

---

📄 Third-party attribution

Bundled OpenMMLab- and TEED-derived components retain their original license notices under licenses/. Detailed third-party attribution is recorded in NOTICE.