mmrotate-SSOOD: Simplified Framework for Semi-Supervised Oriented Object Detection

Introduction

mmrotate-SSOOD is a simplified, modular, and flexible framework designed specifically for Semi-Supervised Oriented Object Detection (SSOOD) tasks.

Support New Method!!!

We now support Denser Teacher(TCSVT 2025)(https://ieeexplore.ieee.org/document/10802941)!

Key Features

Our framework offers the following advantages:

• Simplified Implementation: Implementing custom semi-supervised detection methods is straightforward. You only need to modify two key functions, allowing faster experimentation and development.
• Flexible Data Augmentation: Built upon MMCV, our framework supports seamless integration of custom data augmentation techniques. We also provide ready-to-use augmentation configs for fair comparisons with prior works.
• Dataset Splitting Tools: Easily split your dataset into labeled and unlabeled subsets using our user-friendly tools, saving time on data preparation for semi-supervised learning.
• Extensible Method Support: Currently, we support Denser Teacher(TCSVT 2025)(https://ieeexplore.ieee.org/document/10802941), SOOD(CVPR 2023)(https://arxiv.org/abs/2304.04515) and Dense Teacher(ECCV 2022)(https://arxiv.org/abs/2207.02541), with plans to add more state-of-the-art semi-supervised learning methods in future updates.

Future Methods Support

We plan to continually update this framework to include more state-of-the-art semi-supervised learning methods. Here are some of the methods we aim to support in future updates:

1. Soft Teacher(ICCV 2021) (End-to-End Semi-Supervised Object Detection with Soft Teacher)(https://arxiv.org/abs/2106.09018) A famous semi-supervised object detection method that uses teacher-student architecture with pseudo-label refinement for better performance.

2. ARSL(CVPR 2023) (Ambiguity-Resistant Semi-Supervised Learning for Dense Object Detection)(https://arxiv.org/abs/2303.14960) Focuses on ambiguities in semi-supervised object detection.

Requirements

To ensure compatibility, please install the following dependencies:

1. PyTorch

• PyTorch: 1.13.x
  We recommend PyTorch 1.13.x as all modules have been tested with this version. Installation guide: PyTorch.org

2. MMDetection

• MMDetection: 3.0.0
  MMDetection serves as the base object detection framework. Refer to the MMDetection documentation for installation instructions.

3. MMPretrain

• MMPretrain: 1.1.0
  MMPretrain is used for pretraining models. Please follow the MMPretrain installation guide.

Notes:

• CUDA Compatibility: Make sure all dependencies match your system's CUDA version for proper GPU acceleration. Check the PyTorch documentation for compatibility.
• Virtual Environment: For a cleaner setup, we highly recommend using a virtual environment like conda or venv.

Installation Example:

Here’s a quick guide to set up the environment:

# Create a virtual environment
conda create -n ssood python=3.10
conda activate ssood

# Install PyTorch

# Install mmdet
pip install -U openmim
mim install mmengine
mim install "mmcv==2.0.0"
mim install mmdet==3.0.0

# Install mmpretrain
mim install mmpretrain==1.1.0
pip install future tensorboard
pip install -v -e .

Data Preparation

Please refer to data_preparation.md to prepare the original data. After that, the data folder should be organized as follows:

├── data
│   ├── split_ss_dota1_5
│   │   ├── train
│   │   │   ├── images
│   │   │   ├── annfiles
│   │   ├── val
│   │   │   ├── images
│   │   │   ├── annfiles
│   │   ├── test
│   │   │   ├── images
│   │   │   ├── annfiles

For partial labeled setting, we split the DOTA-v1.5's train set via the author released split data list and split tool:

python tools/SSOD/split_dota1.5_via_lists.py

For fully labeled setting, we use DOTA-V1.5 train as labeled set and DOTA-V1.5 test as unlabeled set.

After that, the data folder should be organized as follows:

├── data
│   ├── split_ss_dota1_5
│   │   ├── train
│   │   │   ├── images
│   │   │   ├── annfiles
│   │   ├── train_10_labeled
│   │   │   ├── images
│   │   │   ├── annfiles
│   │   ├── train_10_unlabeled
│   │   │   ├── images
│   │   │   ├── annfiles
│   │   ├── train_20_labeled
│   │   │   ├── images
│   │   │   ├── annfiles
│   │   ├── train_20_unlabeled
│   │   │   ├── images
│   │   │   ├── annfiles
│   │   ├── train_30_labeled
│   │   │   ├── images
│   │   │   ├── annfiles
│   │   ├── train_30_unlabeled
│   │   │   ├── images
│   │   │   ├── annfiles
│   │   ├── val
│   │   │   ├── images
│   │   │   ├── annfiles
│   │   ├── test
│   │   │   ├── images
│   │   │   ├── annfiles

For DOTAv1.0, the preparation is the same with DOTAv1.5.

Training

For Denser Teacher

• To train Denser Teacher with 10% labeled data, run:

 CUDA_VISIBLE_DEVICES=0,1 PORT=29501 bash ./tools/dist_train.sh configs/rotated_denser_teacher/rotated-denser-teacher_2xb3-180000k_semi-0.1-dotav1.5.py 2

Results

DOTA1.5

SOOD

Backbone	Setting	mAP50	mAP50 in Paper	Mem (GB)	Config
ResNet50 (1024,1024,200)	10%	47.93	48.63	8.45	config
ResNet50 (1024,1024,200)	20%		55.58		config
ResNet50 (1024,1024,200)	30%		59.23		config

Dense Teacher

Backbone	Setting	mAP50	mAP50 in Paper	Mem (GB)	Config
ResNet50 (1024,1024,200)	10%	47.10	-		config
ResNet50 (1024,1024,200)	20%				config
ResNet50 (1024,1024,200)	30%				config

Denser Teacher

Backbone	Setting	mAP50	Mem (GB)	Config
ResNet50 (1024,1024,200)	1%	20.98		config
ResNet50 (1024,1024,200)	5%	43.40		config
ResNet50 (1024,1024,200)	10%	52.05		config
ResNet50 (1024,1024,200)	20%	57.49		config
ResNet50 (1024,1024,200)	30%	60.40		config

DOTAv1.0

Denser Teacher

Backbone	Setting	mAP50	Mem (GB)	Config
ResNet50 (1024,1024,200)	1%	19.45		config
ResNet50 (1024,1024,200)	5%	45.84		config
ResNet50 (1024,1024,200)	10%	52062		config
ResNet50 (1024,1024,200)	20%	59.20		config
ResNet50 (1024,1024,200)	30%	62.82		config

Acknowledgement

This repo is built upon mmrotate. The implementation of SOOD is based on SOOD. The implementation of Dense Teacher is based on Dense Teacher. Thanks for their open source code.