Simplifying Source-Free Domain Adaptation for Object Detection: Effective Self-Training Strategies and Performance Insights

Yan Hao, Florent Forest and Olga Fink. ECCV, 2024.

• pytorch >= 1.12
• detectron2 0.6

simple-SFOD/
├── cityscapes-to-coco-conversion/: Cityscapes dataset preparation tool
├── sim10k-to-coco/: Sim10k dataset preparation tool
├── kitti-to-coco/: KITTI dataset preparation tool
├── configs/: configuration files for detectron2 models
├── convert_pretrained_model/: convert the pytorch pretrained model to detectron2 format
├── daod/
    ├── data/: submodule for data management
        ├── mappers/: submodule for detectron2 dataset mappers
        ├── transforms/: submodule for data transformations
    ├── engine/
        ├── hooks/: submodule for training hooks
        └── trainers/: submodule for trainer classes
    ├── evaluation/: submodule for evaluators
    └── modeling/: submodule for neural net architectures
        ├── meta_arch/: detectron2 meta-architectures
        ├── proposal_generator/: RPN architectures
        └── roi_heads/: RoI heads architectures
    └── config.py: set-up custom configuration
├── README.md: this readme file
├── train_net_mt.py: adaptation script for teacher-student
└── train_net.py: main script for training and evaluation

The following datasets are supported:

• detectron2 built-in datasets
• Cityscapes (instances-only)
• Foggy Cityscapes (instances-only, $\beta \in {0.005, 0.01, 0.02}$)
• Sim10k
• KITTI

To define the root directory containing the datasets, set the environment variable DETECTRON2_DATASETS, e.g. export DETECTRON2_DATASETS=/path/to/datasets/. If not defined, it will be assumed that datasets are in datasets/ inside the project directory.

First, the cityscapes and foggy cityscapes need to be downloaded from the Cityscapes website. Then, extract the archives and modify the folder names to obtain the following structures:

cityscapes
├── gtFine
│   ├── test
│   ├── train
│   └── val
├── gtFine_trainvaltest
│   ├── license.txt
│   └── README
└── leftImg8bit
    ├── test
    ├── train
    └── val

cityscapes_foggy
└── leftImg8bit
    ├── test
    ├── train
    └── val

The conversion of cityscapes and foggy cityscapes into COCO JSON format for object detection (instances only) is performed using the cityscapes-to-coco-conversion tool. It has been slightly modified to remove segmentations and allow to use a suffix to handle fog level (beta). Run following commands to perform the conversion:

python cityscapes-to-coco-conversion/main.py --dataset cityscapes --datadir /path/to/cityscapes --outdir /path/to/cityscapes/annotations

For foggy cityscapes, run this command with the desired "beta" value:

python cityscapes-to-coco-conversion/main.py --dataset cityscapes --datadir /path/to/cityscapes --outdir /path/to/cityscapes_foggy/annotations --file_name_suffix _foggy_beta_0.02

A new directory annotations will be created, containing JSON files with instance annotations for each split (train, val, test).

Run python train_net.py --help for a usage instructions.

For cityscapes:

python train_net.py --num-gpus 1 --config-file configs/faster_rcnn_VGG_cityscapes_source_new.yaml

For cityscapes to cityscapes foggy:

With teacher-student, use the pre-trained model to initialize weights for both teacher and student. Then run:

python train_net_mt.py --num-gpus 1 --config-file configs/faster_rcnn_VGG_cityscapes_foggy_adaptive_teacher_source_free.yaml

With fixed pseudo-labels, firstly get the AdaBN model, then use it to initialize weights. Run:

python train_net.py --num-gpus 1 --config-file configs/faster_rcnn_VGG_cityscapes_foggy_source_wq.yaml

If this work was useful to you, please cite our paper:

@inproceedings{hao_simplifying_2024,
    title = {Simplifying Source-Free Domain Adaptation for Object Detection: Effective Self-Training Strategies and Performance Insights},
    author = {Hao, Yan and Forest, Florent and Fink, Olga},
    booktitle = {ECCV 2024},
    year = {2024},
}