Dual Teacher: A Semi-Supervised Co-Training Framework for Cross-Domain Ship Detection

项目更新（2026-10-01）：已加入独立的 sup2 GIoU 辅助定位实验，以及小背景日志、配对复算护栏和配置生成修复。 新实验默认不启动，原小背景重加权结果保留。配置生成、验收和训练说明见 sup2 定位约束。旧 audit/dev 报告属于历史材料，不作为本轮有效对照。

Introduction

This is the reserch code of the IEEE Transactions on Geoscience and Remote Sensing 2023 paper.

X. Zheng, H. Cui, C. Xu and X. Lu, "Dual Teacher: A Semi-Supervised Co-Training Framework for Cross-Domain Ship Detection," IEEE Transactions Geoscience and Remote Sensing, 2023.

In this code, we explored the Semi-Supervised Cross-Domain Ship Detection (SCSD) task to improve the cross-domain ship detection performance with a few labeled SAR images. We proposed Dual Teacher framework to integrate cross-domain object detection and semi-supervised object detection for different knowledge fusion.

Usage

Requirements

• Ubuntu 20.04
• Anaconda3 with python=3.6
• Pytorch=1.7.0
• mmdetection=2.16.0+fe46ffe
• mmcv=1.3.9
• wandb=0.10.31

Installation

make install

Data

• Download DIOR, HRSID and SSDD datasets and put them as follows:
• Execute the following command to generate data set splits:

# YOUR_DATA/
#   dior/
#     dior_annotations.json    # only ship instances
#     images/
#   hrsid/
#     annotations/
#     images/
#   ssdd/
#     annotations/
#     JPEGImages/
#   dior_hrsid/
#     annotations/             # labeled optical images and few labeled SAR images
#     images/
ln -s ${YOUR_DATA} data
bash tools/dataset/semi_hrsid.sh
bash tools/dataset/semi_ssdd.sh

• ADD HRSIDDataset to MMDetection, similar to COCODataset

Training

# num_SAR_images: number of labeled SAR images for training
# num_gpus: number of gpus for training
bash tools/dist_train_ship_pretrain.sh dior 0 100 ${num_gpus}
for fold in 1,2,3,4,5;
do
    bash tools/dist_train_ship_pretrain.sh dior_hrsid ${fold} ${num_SAR_images} ${num_gpus}
    bash tools/dist_train_dual_teacher_partially_hrsid.sh semi ${fold} ${num_SAR_images} ${num_gpus}
done 

Evaluation

python tools/test.py <config_file_path> <model_file_path> --eval bbox --work-dir <save_dir>

Corrected SSDD baseline: initialization and NMS

The corrected fresh-run path explicitly loads Phase 1 into teacher1/student1 and Phase 2 into teacher2/student2 after generic model initialization and before the runner's first step / EMA hook. Every parameter and buffer is checked for matching keys, shapes, finite values and equality after loading. Missing/incompatible checkpoints stop training. A successful startup prints four [DualTeacher init] ... verified ... messages followed by:

[DualTeacher init] PASS: T1=S1, T2=S2, T1!=T2; fusion=NMS, fusion_iou=0

Pseudo-label fusion now follows the released author code's ordinary NMS, including fusion IoU=0 and empty-teacher passthrough. This is separate from the detector's own NMS thresholds. The previous consensus OR score boost and single-teacher rescaling have been removed. The learning rate, 32000 iterations, loss weights, data splits and EMA schedule of the single-GPU reproduction config are otherwise unchanged; this is not a claim of identical four-GPU optimization.

On the training machine, first check the real checkpoints without a dataset, GPU training or optimizer (repeat for folds 6, 7 and 8):

python tools/check_dual_teacher_init.py configs/reproduce/phase3_dual_teacher_ssdd.py \
    --cfg-options fold=6 percent=3

After the checks pass, launch a new Phase 3 run, retaining the same labeled images. For example (choose and record the training RNG seed deliberately; the fold number only selects the labeled-data split):

python -m torch.distributed.launch --nproc_per_node=1 \
    tools/train_ablation.py configs/reproduce/phase3_dual_teacher_ssdd.py \
    --launcher pytorch --seed 678 --cfg-options fold=6 percent=3

Outputs go to work_dirs/phase3_dual_teacher_baseline_nms/3/6, not the old phase3_dual_teacher directory; automatic resume is disabled. Do not resume the old consensus/uninitialized Phase 3 checkpoints as a corrected baseline. Existing Phase 1/2 checkpoints and all old logs should be kept unchanged. Resume a corrected run only with an explicit --resume-from pointing to its full four-branch checkpoint. Full checkpoint restore/inference does not need the Phase 1/2 files and will not substitute their weights.

CPU regression tests use small synthetic checkpoints and real PyTorch/NMS, with MMDetection construction replaced by lightweight fixtures:

python -m pytest -q tests/test_dual_teacher_baseline.py

These tests require PyTorch, NumPy, Numba and pytest. They do not replace the real-checkpoint startup check above or a full CUDA training experiment.

Fresh isolated ablations after the corrected reproduction

All earlier experimental results are invalid and must not be used as evidence. The current workflow freezes a user-selected correct B0 and generates M2 only, FG only, PG on labeled SAR only, and PG on both SAR branches. Every group inherits B0's dataset, initialization, optimizer and evaluation settings.

The old M1 quality head, M3 routing and MVDT modules and combined dev experiment entry points have been removed; prior versions remain in Git history. Strict Phase1/2 initialization, author-code NMS fusion and the default baseline path remain available. PG is a new unvalidated ablation, not a complete D3T reproduction.

Use the Chinese offline installation and training plan. tools/prepare_ablation_suite.py records configuration differences and input hashes. tools/train_ablation.py pins the actual worker imports, dispatches CPU initialization / real CUDA batch checks, and launches training or evaluation. No training is started by installing this update.

python -m pytest -q tests

CPU tests cover strict initialization, original NMS, M2 weights/normalization, foreground loss/gradients, PG scheduling/resume and configuration isolation. Real MMDetection/CUDA checks must still run on the training machine.

Cite

@article{zheng2023dual,
  author={Zheng, Xiangtao and Cui, Haowen and Xu, Chujie and Lu, Xiaoqiang},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Dual Teacher: A Semisupervised Cotraining Framework for Cross-Domain Ship Detection}, 
  year={2023},
  volume={61},
  number={},
  pages={1-12},
  doi={10.1109/TGRS.2023.3287863}}

Acknowledgement

A large part of the codes are borrowed from SoftTeacher. Thanks for the excellent work!