FlowCLAS: Enhancing Normalizing Flow-Based Anomaly Segmentation Via Contrastive Learning

WACV 2026
Chang Won Lee1, Selina Leveugle1, Paul Grouchy2, Chris Langley2, Svetlana Stolpner2,
Jonathan Kelly1, Steven L. Waslander1
1University of Toronto 2MDA Space
Paper arXiv Video Code coming soon

Code is not public yet. The GitHub repository will be updated here once the release is ready.

🌟 Key Features

  • Hybrid objective for normalizing flows. We introduce FlowCLAS, a framework that combines maximum-likelihood training with an outlier-aware contrastive objective, encouraging a more separable latent space for robust anomaly segmentation in dynamic scenes.
  • Contrastive learning is the key ingredient. Ablations show that contrastive learning outperforms other outlier-based training strategies for normalizing flows and is critical for learning object-level semantic features rather than low-level patterns.
  • Strong performance across multiple domains. FlowCLAS delivers strong results on both road anomaly segmentation benchmarks for autonomous driving (Fishyscapes Lost & Found, Road Anomaly, SegmentMeIfYouCan-ObstacleTrack) and the ALLO benchmark for space anomaly segmentation, showing that the approach transfers well across very different robotics settings.

🧠 Method Overview

FlowCLAS diagram

FlowCLAS Overview. The framework operates in two stages. During training, a frozen vision encoder fφ extracts features from a mixed-content image xmix and an auxiliary outlier image xout. A normalizing flow network fθ then maps these features to a latent space, producing z{mix,out}. The model is optimized via a hybrid objective: (1) a maximum likelihood loss (Lml) pushes the latent samples z corresponding to normal regions toward a base Multivariate Gaussian distribution, and (2) a contrastive loss (Lcon) enforces a separation between the latent representations of normal and anomalous features in the projection space. During inference, the normalizing flow computes a likelihood-based anomaly score map for a given test image, localizing regions that deviate from the learned distribution of normal data.

📊 Results at a Glance

FlowCLAS establishes new state-of-the-art on multiple road and space anomaly segmentation benchmarks while remaining competitive on the most challenging SMIYC split.

Road anomaly segmentation

Fishyscapes Lost & Found

AUPRC 88.8
FPR95 0.7

Best overall result on FS-L&F, improving AUPRC by 3.2 points and reducing FPR95 by 0.6 relative to the strongest prior method reported in the paper.

Road anomaly segmentation

Road Anomaly

AUPRC 93.0
FPR95 3.3

Sets a new best AUPRC and improves FPR95 by 2.0 points over the previous strongest baseline on Road Anomaly.

SegmentMeIfYouCan benchmark

SegmentMeIfYouCan ObstacleTrack

AUPRC 94.2
FPR95 0.1

Achieves the best ObstacleTrack result. On the harder AnomalyTrack split, FlowCLAS reaches 94.3 AUPRC and remains competitive with the top entry.

Space anomaly segmentation

ALLO

AUPRC 88.4
FPR95 6.6

Delivers a 7.6-point AUPRC gain over the previous best result on ALLO while keeping false positives competitive in this more difficult low-light orbital setting.

Why It Works

The paper's ablations show that contrastive learning is the decisive ingredient rather than a minor training detail.

Contrastive learning beats other OE strategies

Outlier exposure helps, but the strongest results come from enforcing latent-space separation with a contrastive loss instead of relying only on segmentation-style supervision or outlier likelihood minimization.

The objective transfers to existing NF models

On ALLO, adding the FlowCLAS contrastive objective improves FastFlow from 29.1 to 40.8 AUPRC and UFlow from 20.7 to 48.7 AUPRC, showing the framework is a reusable upgrade rather than a one-off architecture.

Strong features help, but they are not enough alone

Better pre-trained encoders improve the ceiling, yet the paper shows that rich features still need the contrastive objective to avoid low-level shortcuts and achieve reliable object-level anomaly separation.

🖼️ Visualizations

Space anomaly segmentation

ALLO test set, sample 1
ALLO qualitative result on a held-out orbital scene. The visualization highlights FlowCLAS's ability to localize an anomalous object while preserving the surrounding ISS structure.
ALLO test set, sample 2
ALLO qualitative result under a second viewpoint and lighting condition, illustrating the method's transfer to dynamic space robotics imagery.

Road anomaly segmentation

Road anomaly segmentation from the Fishyscapes Lost and Found dataset
Fishyscapes Lost & Found validation example showing dense anomaly localization in a cluttered road scene with small hazardous objects.
Road anomaly segmentation from the Road Anomaly dataset
Road Anomaly validation example illustrating object-level segmentation in a diverse urban scene with strong appearance variation.

Citation

@inproceedings{leeFlowCLAS2026,
  author    = {Lee, Chang Won and Leveugle, Selina and Grouchy, Paul and Langley, Chris and Stolpner, Svetlana and Kelly, Jonathan and Waslander, Steven L.},
  title     = {FlowCLAS: Enhancing Normalizing Flow-Based Anomaly Segmentation Via Contrastive Learning},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2026},
}

Acknowledgements

This website is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.

Thank you to the authors of Nerfies for the website template.