[WACV 2026] FlowCLAS: Enhancing Normalizing Flow-Based Anomaly Segmentation Via Contrastive Learning

Official repository for FlowCLAS.

All commands below assume the working directory is src/ (where trainer_cli.py lives).

Environment setup

Docker

The image is defined under docker/. It uses a multi-stage build (CUDA 12.8, PyTorch 2.7, MMCV, SAM2, and project Python dependencies). GPU access at runtime requires the NVIDIA Container Toolkit.

Build (from the repository root):

bash scripts/docker_build.bash

Equivalent manual command (paths are relative to the repo root; the build context must be the docker/ directory):

RENDER_GID=$(getent group render | cut -d: -f3)

docker build \
  --network=host \
  --build-arg "USERNAME=$(whoami)" \
  --build-arg CREATE_USER=true \
  --build-arg "WORKDIR_PATH=/home/$(whoami)" \
  --build-arg "RENDER_GID=${RENDER_GID}" \
  -t flowclas \
  -f docker/Dockerfile \
  docker

Flag / argument	Purpose
--network=host	Uses the host network during build (helps reach package indexes).
CREATE_USER=true	Creates a container user matching your host username (for volume permissions).
WORKDIR_PATH	Sets the container working directory to your home path inside the image.
RENDER_GID	Adds the container user to a render group with the host’s GID (for GPU/GL access).

The image is tagged flowclas. To start an interactive container with project directories mounted, run from the repository root:

bash scripts/docker_run.bash /path/to/your/datasets

Paths in the YAML configs below are relative to src/ (for example, ../data/cityscapes → ~/data/cityscapes inside the container).

Dataset preparation

Cityscapes

1. Register and download the gtFine split (and matching leftImg8bit images) from the Cityscapes dataset.
2. (Recommended) Preprocess labels to the common 19-class trainId convention using cityscapesScripts — install the package, then run createTrainIdLabelImgs.py on your download. FlowCLAS remaps raw labelIds to 19 training classes in code; generating *_gtFine_labelTrainIds.png keeps your tree aligned with the standard Cityscapes tooling.
3. Set city_root in src/configs/base/cityscapes_coco.yaml to the dataset root that contains leftImg8bit/ and gtFine/ (Torchvision layout).

Expected layout:

cityscapes/
├── leftImg8bit/
│   ├── train/<city>/*.png
│   └── val/<city>/*.png
└── gtFine/
    ├── train/<city>/*_gtFine_labelIds.png
    └── val/<city>/*_gtFine_labelIds.png

ALLO

A new version of the ALLO dataset will be released soon. For the current benchmark, set allo_root, train_dir, and test_dir in src/configs/base/allo_coco.yaml.

COCO (outlier exposure)

COCO images are used as an OoD proxy during training (paste/mix with in-distribution scenes). You can either prepare them yourself or use the preprocessed archive on Google Drive (same folder as Fishyscapes and Road Anomaly).

Option A — prepare from scratch (Meta-OoD):

1. Download COCO 2017 train images and instance annotations.
2. Run preparation/prepare_coco_segmentation.py to build binary ood_seg masks for images without instances that overlap Cityscapes train classes.

Option B — download preprocessed data from the Google Drive folder and extract under your coco_root.

In both cases, set coco_root in the data config you use (cityscapes_coco.yaml and/or allo_coco.yaml).

Expected layout:

coco/
├── annotations/
│   ├── instances_train2017.json
│   └── ood_seg_train2017/          # optional; masks may also live at repo root
├── ood_seg_train2017/*.png         # required by FlowCLAS loaders
└── train2017/*.jpg

You can override any root on the CLI, for example:

--data.init_args.city_root /path/to/cityscapes \
--data.init_args.coco_root /path/to/coco

Other benchmarks (preprocessed)

Preprocessed Fishyscapes, Road Anomaly, and COCO (see above) are available on Google Drive. After extracting, point the paths in cityscapes_coco.yaml:

Config key	Role
fishy_root	Fishyscapes Lost & Found split (images/, labels/)
roadanomaly_root	Road Anomaly frames and semantic labels
smiyc_root	Segment Me If You Can (AnomalyTrack + ObstacleTrack)

Expected layouts (as used by this repo):

fishyscapes/
├── LostAndFound/
│   ├── images/*.png
│   └── labels/*.png
└── Static/
    ├── images/*.png
    └── labels/*.png

RoadAnomaly/
└── frames/
    ├── <clip>.jpg
    └── <clip>.labels/
        └── labels_semantic.png

smiyc/
├── dataset_AnomalyTrack/
│   ├── images/
│   └── labels_masks/
└── dataset_ObstacleTrack/
    ├── images/
    └── labels_masks/

Evaluation uses fishy_root for Fishyscapes (default: ../data/fishyscapes/LostAndFound), plus smiyc_root and roadanomaly_root for the other test benchmarks in the Cityscapes training pipeline.

Weights and checkpoints

Pre-trained backbones, SAM2, and FlowCLAS checkpoints are available on Google Drive. After downloading, place files under the repo root as follows (paths are relative to src/ in the configs):

File	Destination
rein_dinov2l_allo.pth	weights/rein_dinov2l_allo.pth
rein_dinov2l_city.pth	weights/rein_dinov2l_city.pth
sam2.1_hiera_large.pt	misc/weights/sam2.1_hiera_large.pt
best_allo.ckpt	weights/flowclas/best_allo.ckpt
best_cityscapes.ckpt	weights/flowclas/best_cityscapes.ckpt

Training reads backbone_ckpt from flowclas_allo.yaml / flowclas_city.yaml; inference passes --ckpt to a FlowCLAS checkpoint under weights/flowclas/.

Results

Pre-trained checkpoints and pixel-level metrics on the official test splits (checkpoints from Google Drive):

Benchmark	Data config	Model config	Checkpoint	Pixel AP	Pixel FPR95
ALLO	config	config	best_allo.ckpt	88.4	6.6
Fishyscapes	config	config	best_cityscapes.ckpt	88.8	0.7
Road Anomaly	config	config	best_cityscapes.ckpt	93.0	3.3

Fishyscapes and Road Anomaly are evaluated with the same Cityscapes-trained best_cityscapes.ckpt; use the inference commands in Road Anomaly Segmentation below.

Training

Replace <num_devices>, <num_workers>, and <seed> with values for your setup.

ALLO

python3 trainer_cli.py fit \
  --data configs/base/allo_coco.yaml \
  -c configs/base/base.yaml \
  --trainer.devices <num_devices> \
  --data.num_workers <num_workers> \
  -c configs/flowclas/flowclas_allo.yaml \
  --seed_everything <seed> \
  --experiment.name flowclas_allo \
  --experiment.project_name ALLO

Cityscapes (Road Anomaly / Fishyscapes)

python3 trainer_cli.py fit \
  --data configs/base/cityscapes_coco.yaml \
  -c configs/base/base.yaml \
  --trainer.devices <num_devices> \
  --data.num_workers <num_workers> \
  -c configs/flowclas/flowclas_city.yaml \
  --seed_everything <seed> \
  --experiment.name flowclas_city \
  --metrics_on_cpu \
  --no_sigmoid

Examples

Paper-style runs using alternate experiment configs (same flag set as above):

ALLO

python3 trainer_cli.py fit \
  --data configs/base/allo_coco.yaml \
  -c configs/base/base.yaml \
  --trainer.devices 4 \
  --data.num_workers 8 \
  -c configs/flowclas/flowclas_allo.yaml \
  --seed_everything 42 \
  --experiment.name flowclas_allo \
  --experiment.project_name flowclas_ALLO

Cityscapes

python3 trainer_cli.py fit \
  --data configs/base/cityscapes_coco.yaml \
  -c configs/base/base.yaml \
  --trainer.devices 2 \
  --data.num_workers 8 \
  -c configs/flowclas/flowclas_city.yaml \
  --seed_everything 42 \
  --experiment.name flowclas_city \
  --experiment.project_name flowclas_Cityscapes \
  --metrics_on_cpu \
  --no_sigmoid

Inference

Replace <num_devices>, <num_workers>, <seed>, and <ckpt> with values for your setup.

ALLO

python3 trainer_cli.py test \
  --data configs/base/allo_coco.yaml \
  -c configs/base/base.yaml \
  --trainer.devices <num_devices> \
  --data.num_workers <num_workers> \
  -c configs/flowclas/flowclas_allo.yaml \
  --seed_everything <seed> \
  --experiment.name flowclas_allo \
  --data.eval_batch_size 2 \
  --experiment.logger [csv] \
  --trainer.precision 32 \
  --ckpt <ckpt>

Road Anomaly Segmentation (Fishyscapes & Road Anomaly)

python3 trainer_cli.py test \
  --data configs/base/cityscapes_coco.yaml \
  -c configs/base/base.yaml \
  --trainer.devices <num_devices> \
  --data.num_workers <num_workers> \
  -c configs/flowclas/flowclas_city.yaml \
  --seed_everything <seed> \
  --experiment.name flowclas_city \
  --data.eval_batch_size 2 \
  --metrics_on_cpu \
  --no_sigmoid \
  --experiment.logger [csv] \
  --trainer.precision 32 \
  --ckpt <ckpt>

Examples

Download checkpoints from Google Drive first (see Weights and checkpoints). Example inference commands (same flag set as above):

ALLO

python3 trainer_cli.py test \
  --data configs/base/allo_coco.yaml \
  -c configs/base/base.yaml \
  --trainer.devices 1 \
  --data.num_workers 8 \
  -c configs/flowclas/flowclas_allo.yaml \
  --seed_everything 42 \
  --experiment.name flowclas_allo_test \
  --data.eval_batch_size 2 \
  --experiment.logger [csv] \
  --trainer.precision 32 \
  --ckpt ../weights/flowclas/best_allo.ckpt

Road Anomaly Segmentation (Fishyscapes & Road Anomaly)

python3 trainer_cli.py test \
  --data configs/base/cityscapes_coco.yaml \
  -c configs/base/base.yaml \
  --trainer.devices 1 \
  --data.num_workers 1 \
  -c configs/flowclas/flowclas_city.yaml \
  --seed_everything 42 \
  --experiment.name flowclas_city_test \
  --data.eval_batch_size 2 \
  --metrics_on_cpu \
  --no_sigmoid \
  --experiment.logger [csv] \
  --trainer.precision 32 \
  --ckpt ../weights/flowclas/best_cityscapes.ckpt

Citation

If you use FlowCLAS in your research, please cite:

@inproceedings{lee2026flowclas,
  title={FlowCLAS: Enhancing Normalizing Flow-Based Anomaly Segmentation Via Contrastive Learning},
  author={Lee, Chang Won and Leveugle, Selina and Grouchy, Paul and Langley, Chris and Stolpner, Svetlana and Kelly, Jonathan and Waslander, Steven L},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages={6998--7007},
  year={2026}
}