✨ Enhancing Cross-domain Few-annotation Object Detection via Memory Storage-to-Adaptation Mechanism [Project Page]

This work was accepted by The Visual Computer (2026-05-04). This repository contains the code for reproducing the paper: Enhancing Cross-domain Few-annotation Object Detection via Memory Storage-to-Adaptation Mechanism

Overview

• We propose a novel memory storage-to-adaptation mechanism to learn the prior knowledge and transfer it to feature alignment adaptively. To the best of our knowledge, this is the first work to extract the prior knowledge of unlabeled target data to address the CFOD task.
• We construct a new challenging benchmark of industrial scenarios, that comprises complex background and large differences between source and target domains.
• Experiments show that the proposed MS2A outperforms state-of-the-art methods on both public datasets and the constructed industrial dataset. In particular, MS2A exceeds the state-of-the-art method by 10.4% on the challenging industrial dataset for the 10-annotation setting.

Example Results

Qualitative comparisons on public datasets C $\rightarrow$ F setting.

Qualitative comparisons on public datasets K $\rightarrow$ C setting.

Qualitative comparisons on proposed industrial datasets Indus_S $\rightarrow$ Indus_T1 setting.

Qualitative comparisons on proposed industrial datasets Indus_S $\rightarrow$ Indus_T2 setting.

Installation

1、create conda environment and activate it

conda create -n MS2A python=3.9
conda activate MS2A

2、install torch、torchvision、mmcv and cmake

pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113
pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu113/torch1.11.0/index.html
conda install cmake

3、git clone MS2A project and run the following command

git clone https://github.com/GuHuangAI/MS2A.git
cd MS2A
pip install -r requirements.txt
pip install -v -e .

Reference to Installation for more installation instructions.

Dataset

Please download our proposed Industrial Datasets from here. It consists of two settings, Indus_S $\rightarrow$ Indus_T1 and Indus_S $\rightarrow$ Indus_T2. We use LabemMe for labeling and annotate all the objects as the part class. More information about the size of the training set and test set can be obtained by browsing the table below.

	Indus_S→Indus_T1		Indus_S→Indus_T2
	Source	Target	Source	Target
train	4614	10、30、50	4614	10、30、50
test	1153	269	1153	432

Pretrained model for quick start

All used pretrained models can be downloaded from here.The yolox_x_8x8_300e_coco_20211126_140254-1ef88d67.pth is the pretrained model from mmdetection. city_foggycity_epoch_20.pth, kitti_city_epoch_18.pth and sim10k_city_epoch_15.pth are pretrained model on public dataset cityscapes2foggycityscapes, kitti2cityscapes and sim10k2cityscapes task respectively.

Training from scratch

Stage I - Memory Storage

1. pretrain base detector
./tools/dist_train.sh ${config_file} ${gpu_number}
e.g: ./tools/dist_train.sh ./configs/yolox/yolox_fs_x_640x640_50e_18_base.py 1

2. extract and cluster the prior knowledge
python extract_fea_fewshot.py \
 --depart ${dataset name} \
 --img ${image dir} \
 --config ${config file} \
 --checkpoint ${checkpoint file} \
 --out_path ${save dir} \
 --batch_size 8 \
 --dim ${128/256 large:256}
e.g: 
python extract_fea_fewshot.py \
 --depart Indus_T1 \
 --img /nas/public_data/foggy_cityscapes_COCO_format/train2014 \
 --config ./configs/yolox/yolox_fs_x_640x640_50e_cityscapes_base.py \
 --checkpoint ./work_dirs/yolox_fs_x_640x640_50e_cityscapes_base/best_bbox_mAP_epoch_70.pth \
 --out_path /path/to/save/extracted/feture \
 --batch_size 8 \
 --dim 320

python knn_cuda.py \
 --fea_root_path ${last step saved dir} \
 --out_path ${default same as fea_root_path} \
 --n_clusters ${default 100} \
 --depart ${cluster name}
e.g:
python knn_cuda.py \
 --fea_root_path /path/to/save/extracted/feture/ \
 --out_path /path/to/save/cluster/results/ \
 --n_clusters 100 \
 --depart Indus_T1

Stage II - Memory Adaptation

1. Source Training
./tools/dist_train.sh ${config_file} ${gpu_number}
e.g: ./tools/dist_train.sh configs/yolox/yolox_fs_x_640x640_50e_cityscapes_pretrain_mom.py 1

2. Target Training
./tools/dist_train.sh ${config_file} ${gpu_number}
e.g: ./tools/dist_train.sh configs/yolox/yolox_fs_x_640x640_50e_cityscapes_pretrain_mom_ft.py 1

Evaluation

python tools/test.py ${config file} ${checkpoint file} [--out ${RESULT_FILE}] [--eval ${EVAL_METRICS}] [--show]
e.g:
python tools/test.py \
 ./configs/yolox/yolox_fs_x_640x640_50e_cityscapes_pretrain_mom_ft.py \
 ./work_dirs/yolox_fs_x_640x640_50e_cityscapes_pretrain_mom_ft/best_0_bbox_mAP_epoch_20.pth \
 --eval bbox

Inference

python tools/inference.py ${config file} ${checkpoint file} ${save dir}
e.g:
python tools/inference.py \
 ./configs/yolox/yolox_fs_x_640x640_50e_cityscapes_pretrain_mom_ft.py \
 ./work_dirs/yolox_fs_x_640x640_50e_cityscapes_pretrain_mom_ft/best_0_bbox_mAP_epoch_20.pth \
 ./path/to/save

Thanks

Thanks to the public repos: mmdetection for providing the base code.