This repository provides the official code for the WACV 2023 paper Refign: Align and Refine for Adaptation of Semantic Segmentation to Adverse Conditions. The code is organized using PyTorch Lightning.
π₯ [September 2, 2022] Applied on top of HRDA, Refign ranks #1 on both the ACDC leaderboardβ72.05 mIoUβand the Dark Zurich leaderboardβ63.91 mIoU. See below for training configurations.
Due to the scarcity of dense pixel-level semantic annotations for images recorded in adverse visual conditions, there has been a keen interest in unsupervised domain adaptation (UDA) for the semantic segmentation of such images. UDA adapts models trained on normal conditions to the target adverse-condition domains. Meanwhile, multiple datasets with driving scenes provide corresponding images of the same scenes across multiple conditions, which can serve as a form of weak supervision for domain adaptation. We propose Refign, a generic extension to self-training-based UDA methods which leverages these cross-domain correspondences. Refign consists of two steps: (1) aligning the normal-condition image to the corresponding adverse-condition image using an uncertainty-aware dense matching network, and (2) refining the adverse prediction with the normal prediction using an adaptive label correction mechanism. We design custom modules to streamline both steps and set the new state of the art for domain-adaptive semantic segmentation on several adverse-condition benchmarks, including ACDC and Dark Zurich. The approach introduces no extra training parameters, minimal computational overheadβduring training onlyβand can be used as a drop-in extension to improve any given self-training-based UDA method.
The code is run with Python 3.8.13. To install the packages, use:
pip install -r requirements.txt
The following environment variable must be set:
export DATA_DIR=/path/to/data/dir
Before running the code, download and extract the corresponding datasets to the directory $DATA_DIR
.
Cityscapes
Download leftImg8bit_trainvaltest.zip and gt_trainvaltest.zip from here and extract them to $DATA_DIR/Cityscapes
.
$DATA_DIR
βββ Cityscapes
β βββ leftImg8bit
β β βββ train
β β βββ val
β βββ gtFine
β β βββ train
β β βββ val
βββ ...
Afterwards, run the preparation script:
python tools/convert_cityscapes.py $DATA_DIR/Cityscapes
ACDC
Download rgb_anon_trainvaltest.zip and gt_trainval.zip from here and extract them to $DATA_DIR/ACDC
.
$DATA_DIR
βββ ACDC
β βββ rgb_anon
β β βββ fog
β β βββ night
β β βββ rain
β β βββ snow
β βββ gt
β β βββ fog
β β βββ night
β β βββ rain
β β βββ snow
βββ ...
Dark Zurich
Download Dark_Zurich_train_anon.zip, Dark_Zurich_val_anon.zip, and Dark_Zurich_test_anon_withoutGt.zip from here and extract them to $DATA_DIR/DarkZurich
.
$DATA_DIR
βββ DarkZurich
β βββ rgb_anon
β β βββ train
β β βββ val
β β βββ val_ref
β β βββ test
β β βββ test_ref
β βββ gt
β β βββ val
βββ ...
Nighttime Driving
Download NighttimeDrivingTest.zip from here and extract it to $DATA_DIR/NighttimeDrivingTest
.
$DATA_DIR
βββ NighttimeDrivingTest
β βββ leftImg8bit
β β βββ test
β βββ gtCoarse_daytime_trainvaltest
β β βββ test
βββ ...
BDD100k-night
Download 10k Images
and Segmentation
from here and extract them to $DATA_DIR/bdd100k
.
$DATA_DIR
βββ bdd100k
β βββ images
β β βββ 10k
β βββ labels
β β βββ sem_seg
βββ ...
RobotCar for Segmentation
Download all data from here and save them to $DATA_DIR/RobotCar
. As mentioned in the corresponding README.txt, the images must be downloaded from this link.
$DATA_DIR
βββ RobotCar
β βββ images
β β βββ dawn
β β βββ dusk
β β βββ night
β β βββ night-rain
β β βββ ...
β βββ correspondence_data
β β βββ ...
β βββ segmented_images
β β βββ training
β β βββ validation
β β βββ testing
βββ ...
MegaDepth
For training, we use the version provided by the D2-Net repo. Follow their instructions for downloading and preprocessing the dataset.
For testing, we use the split provided by RANSAC-Flow here.
The directories MegaDepth_Train
, MegaDepth_Train_Org
, and Val
can be removed.
All in all, the folder structure should look as follows:
$DATA_DIR
βββ MegaDepth
β βββ Undistorted_SfM
β β βββ ...
β βββ scene_info
β β βββ ...
β βββ Test
β β βββ test1600Pairs
β β | βββ ...
β β βββ test1600Pairs.csv
βββ ...
RobotCar for Matching
We use the correspondence file provided by RANSAC-Flow here. If not already downloaded for segmentation, download the images from here.
$DATA_DIR
βββ RobotCar
β βββ images
β β βββ dawn
β β βββ dusk
β β βββ night
β β βββ night-rain
β β βββ ...
β βββ test6511.csv
βββ ...
We provide pretrained models of both UDA and alignment networks.
Note that the UAWarpC checkpoint is needed to train Refign. To avoid config file edits, save it to ./pretrained_models/
.
To facilitate qualitative segmentation comparisons, validation set predictions of Refign can be directly downloaded. Starred models use Cityscapes pretrained weights in the backbone, the others ImageNet pretrained.
Model | Task | Test Set | Test Score | Config | Checkpoint | Predictions |
---|---|---|---|---|---|---|
Refign-DAFormer | CityscapesβACDC | ACDC test | 65.5 mIoU | config | model | ACDC val |
Refign-HRDA* | CityscapesβACDC | ACDC test | 72.1 mIoU | config | model | ACDC val |
Refign-DAFormer | CityscapesβDark Zurich | Dark Zurich test | 56.2 mIoU | config | model | Dark Zurich val |
Refign-HRDA* | CityscapesβDark Zurich | Dark Zurich test | 63.9 mIoU | config | model | Dark Zurich val |
Refign-DAFormer | CityscapesβRobotCar | RobotCar Seg. test | 60.5 mIoU | config | model | RobotCar val |
Model | Task | Test Set | Score | Config | Checkpoint |
---|---|---|---|---|---|
UAWarpC | MegaDepth Dense Matching | RobotCar Matching test | 36.8 PCK-5 | stage1, stage2 | model |
Make sure to first download the trained UAWarpC model with the link provided above.
Enter the path to the UAWarpC model for model.init_args.alignment_head.init_args.pretrained
in the config file you intend to run (or save the model to ./pretrained_models/
).
To train Refign on ACDC (single GPU, with AMP) use the following command:
python tools/run.py fit --config configs/cityscapes_acdc/refign_hrda_star.yaml --trainer.gpus 1 --trainer.precision 16
Similar config files are available for Dark Zurich and RobotCar.
We also provide the config files for reproducing the ablation study in configs/cityscapes_acdc/ablations/
.
As mentioned in the previous section, modify the config file by adding the UAWarpC model path. To evaluate Refign e.g. on the ACDC validation set, use the following command:
python tools/run.py test --config configs/cityscapes_acdc/refign_hrda_star.yaml --ckpt_path /path/to/trained/model --trainer.gpus 1
We also provide pretrained models, which can be downloaded from the link above. To evaluate them, simply provide them as the argument --ckpt_path
.
To get test set scores for ACDC and DarkZurich, predictions are evaluated on the respective evaluation servers: ACDC and DarkZurich. To create and save test predictions for e.g. ACDC, use this command:
python tools/run.py predict --config configs/cityscapes_acdc/refign_hrda_star.yaml --ckpt_path /path/to/trained/model --trainer.gpus 1
Alignment training consists of two stages. To train stage 1 use:
python tools/run.py fit --config configs/megadepth/uawarpc_stage1.yaml --trainer.gpus 1 --trainer.precision 16
Afterwards, launch stage 2, providing the path of the last checkpoint of stage 1 as an argument:
python tools/run.py fit --config configs/megadepth/uawarpc_stage2.yaml --model.init_args.pretrained /path/to/last/ckpt/of/stage1 --trainer.gpus 1 --trainer.precision 16
Training of the alignment network takes several days on a single GPU.
We use a separate config file for evaluation, to avoid the lengthy sampling of MegaDepth training data at that stage. But of course, the config file used for training can be used as well.
python tools/run.py test --config configs/megadepth/uawarpc_evalonly.yaml --ckpt_path /path/to/last/ckpt/of/stage2 --trainer.gpus 1
We also provide a pretrained model, which can be downloaded from the link above. To test it, simply provide it as the argument --ckpt_path
.
Local correlation is implemented through this custom CUDA extension. By default the extension is built just in time. In case there are problems with this mechanism, the extension can be alternatively pre-installed in the environment (see also the README of the linked repo):
pip install spatial-correlation-sampler
Check the training_step
method in models/segmentation_model.py
. You will need to implement similar logic as is called when the use_refign
attribute is True
. In particular, you also need methods align
and refine
, located in the same file (and the full alignment network). Of course, the dataloader must also return a reference image for Refign to work.
If you find this code useful in your research, please consider citing the paper:
@inproceedings{bruggemann2022refign,
title={Refign: Align and Refine for Adaptation of Semantic Segmentation to Adverse Conditions},
author={Bruggemann, David and Sakaridis, Christos and Truong, Prune and Van Gool, Luc},
booktitle={WACV},
year={2023}
}
This repository is released under the MIT license. However, care should be taken to adopt appropriate licensing for third-party code in this repository. Third-party code is marked accordingly.
For questions about the code or paper, feel free to contact me (send email).