Similarity Min-Max: Zero-Shot Day-Night Domain Adaptation

ICCV 2023 (Oral Presentation)

---

This the official repository of the paper Similarity Min-Max: Zero-Shot Day-Night Domain Adaptation.

For more information, please visit our project website.

Authors: Rundong Luo¹, Wenjing Wang¹, Wenhan Yang², Jiaying Liu¹

¹Peking University, ²Peng Cheng Laboratory

Abstract

Low-light conditions not only hamper human visual experience but also degrade the model's performance on downstream vision tasks. While existing works make remarkable progress on day-night domain adaptation, they rely heavily on domain knowledge derived from the task-specific nighttime dataset. This paper challenges a more complicated scenario with border applicability, i.e., zero-shot day-night domain adaptation, which eliminates reliance on any nighttime data. Unlike prior zero-shot adaptation approaches emphasizing either image-level translation or model-level adaptation, we propose a similarity min-max paradigm that considers them under a unified framework. On the image level, we darken images towards minimum feature similarity to enlarge the domain gap. Then on the model level, we maximize the feature similarity between the darkened images and their normal-light counterparts for better model adaptation. To the best of our knowledge, this work represents the pioneering effort in jointly optimizing both aspects, resulting in a significant improvement of model generalizability. Extensive experiments demonstrate our method's effectiveness and broad applicability on various nighttime vision tasks, including classification, semantic segmentation, visual place recognition, and video action recognition.

Updates

• 09/06/2023: Code for image classification and semantic segmentation is available now.
• 10/18/2023: Code for visual place recognition is available now.
• 11/25/2023: Code for video action recognition is available now.

Code

Code for image classification and semantic segmentation is available now. Code for visual place recognition and video action recognition will be released soon.

Requirements

Environment: Pytorch with version >= 1.11.0 is required. Other requirements can be easily satisfied using pip install.

GPU: 3 GPUs with at least 12GB memory (e.g., 2080Ti) are required.

Image Classification

Dataset Preparation

Download the CODaN dataset and put it under ./classification/data/.

Training

• Navigate to ./darkening, run python darken_classification.py --sim --experiment EXPERIMENT_NAME to train the darkening model with the $\mathcal{L}_D^{sim}$. Specify the pre-trained daytime model path with --sim_model_dir. Specify the logging directory by --experiment, and models will be saved under ./darkening/checkpoints/{args.experiment}.
• Navigate to ./classification, run python train.py --use_BYOL using the BYOL loss ($\mathcal{L}_F^{sim}$). Specify the pre-trained darkening model path with --darkening_model, the pre-trained daytime model with --checkpoint, and the logging directory by --experiment. Model checkpoints and loggers will be saved under ./classification/checkpoints/{args.experiment}.
• Classification results will be saved in ./classification/results/{args.experiment}/log.txt. You may also run python test.py to test the model.

Semantic Segmentation

Dataset Preparation

We need three dataset for training and evaluation: Cityscapes, Nighttime Driving, and Dark Zurich. Download and put them under ./segmentation/data/.

Training

• Navigate to ./darkening, run python darken_segmentation.py --sim --experiment EXPERIMENT_NAME to train the darkening model with the $\mathcal{L}_D^{sim}$. Specify the pre-trained daytime model path with --sim_model_dir, the logging directory by --experiment, and models will be saved under ./darkening/checkpoints/{args.experiment}.
• To save GPU memory, our implementation generates the darkened nighttime dataset in advance. Run python darken_test.py and specify the source daytime dataset path with --src_path, the darkening model path with --experiment, and the target nighttime dataset path with --output_dir. The darkened nighttime dataset will be saved under --dst_path. You may also download our pre-generated darkened nighttime dataset here.
• Navigate to ./segmentation, run python train.py. Specify the darkened dataset by --darken_dataset and the logging directory by --experiment. Model checkpoints and loggers will be saved under ./segmentation/runs/{args.experiment}.
• Segmentation results will be saved in ./segmentation/runs/{args.experiment}/logs/. You may also run python eval_test.py to obtain the visualization results and the zipped file for Dark Zurich evaluation.

Visual Place Recognition

Dataset Preparation

• Please modify the return value of the function get_data_root in ./visual-place-recognition/cirtorch/utils/general.py to the path of the dataset. The datasets will be automatically downloaded and extracted. You may also download the dataset from here and put it under ./visual-place-recognition/dataset/. The training data will be stored in ./retrieval/train/retrieval-SfM-120k.

Training

• Navigate to ./darkening run python darken_vpr.py to train the darkening model with the $\mathcal{L}_D^{sim}$. Specify the pre-trained daytime model path with --sim_model_dir, the logging directory by --experiment, and models will be saved under ./darkening/checkpoints/{args.experiment}.
• Navigate to ./visual-place-recognition, run python3 -m cirtorch.examples.train_night EXPERIMENT_NAME --training-dataset 'retrieval-SfM-120k' --test-datasets '247tokyo1k' --arch 'resnet101' --pool 'gem' --loss 'contrastive' --loss-margin 0.85 --optimizer 'adam' --lr 5e-7 --neg-num 5 --query-size=22000 --pool-size=2000 --batch-size 5 --image-size 362 --epochs 5 --darkening_model PATH_TO_DARKENING_MODEL --sim.
• Model checkpoints and loggers will be saved under ./visual-place-recognition/checkpoints/EXPERIMENT_NAME. Run python3 -m cirtorch.examples.test --network-path PATH_TO_CHECKPOINT --datasets '247tokyo1k' --whitening 'retrieval-SfM-120k' --multiscale '[1, 1/2**(1/2), 1/2]' to test the model.

Low-Light Action Recognition

Dataset Preparation

• Download the normal light data from here (CVPR'22 UG2 challenge). It contains 2,625 videos from 11 classes. Put it under ./low-light-action-recognition/dataset/NormalLight.
• Download the low light data (ARID dataset) from here. Put it under ./low-light-action-recognition/dataset/ARID.

Training

• Navigate to ./low-light-action-recognition, run python train_darkening.py --sim --experiment EXPERIMENT_NAME to train the darkening model with the $\mathcal{L}_D^{sim}$. Specify the pre-trained daytime model path with --feature_extractor (download from our pretrained model link), the logging directory by --experiment, and models will be saved under ./low-light-action-recognition/checkpoints/{args.experiment}.
• Run darken_video.py --darkening_model PATH_TO_DARKENING_MODEL --pretrained_model_path PATH_TO_DAYTIME_PRETRAINED_MODEL. The darkened videos will be saved under the specified directory (use ./dataset/NormalLight/raw/data_darken_test/ by default, otherwise you should also change the darkened data path under ./data/iterator_factory).
• Run train_night.py --model-dir SAVE_DIR --load_checkpoint PATH_TO_PRETRAINED_DAYTIME_MODEL. Model checkpoints and loggers will be saved under ./low-light-action-recognition/checkpoints/{args.model_dir}.
• Run test.py --model-dir PATH_TO_CHECKPOINT for evaluation.

Pre-trained Models

We provide all the data and pre-trained models here.

Citation

If you find this work useful in your research, please consider citing:

@inproceedings{luo2023similarity,
  title={Similarity Min-Max: Zero-Shot Day-Night Domain Adaptation},
  author={Luo, Rundong and Wang, Wenjing and Yang, Wenhan and Liu, Jiaying},
  booktitle={ICCV},
  year={2023}
}

Acknowledgement

Some code are borrowed from CIConv. If you have any questions, please contact Rundong Luo (rundongluo2002@gmail.com) or open an issue.