Contrubute to Anomalib as a potential GSOC 2026 Contributor

[rajeshgangireddy]

Hello All users and contributors of Anomalib,

As usual, Anomalib/OpenVINO is a part of Google Summer of Code (GSoC) 2026, and we’re starting to collect interest from potential contributors.

Suggested Topics:

• Synthetic Anomaly Generation (details), which focuses on creating synthetic defects/anomalies to improve training data for validation, especially in scenarios where real anomalies are rare or expensive to collect.
• Multi-Camera Support for Anomalib (details) where the objective is to extend the "Anomalib Studio" application to support multi-camera streams.

These ideas are just a few suggestions, so contributors are welcome to propose other interesting project topics that fit Anomalib. Also, there may be other projects (outside Anomalib) that you might find exciting. Please take a look at the full list of suggest topics from Intel OpenVINO and Intel Open Edge organisation here: https://github.com/openvinotoolkit/openvino/wiki/Project-ideas-for-2026.

Questions? Feel free to start a discussion here or
simple comment below this issue as well.

Disclaimer: project ideas are not final and will only be confirmed after the GSoC organization/application process approves them.

[github-actions]

Thanks for submitting this issue! It has been added to our triage queue. A maintainer will review it shortly.

[AbhayKumarDas]

Hi @maintainers (cc: @RajkumarReddy and @ashwinvaidya17),

I have submitted my GSoC 2026 proposal on improving synthetic anomaly generation for threshold calibration in Anomalib and wanted to share a quick summary with results. While working with PerlinAnomalyGenerator + synthetic.py, I observed that anomalies are generated with very similar difficulty (blend_factor ~0.01–0.2 and hard binary masks), which leads to tightly clustered anomaly scores and a gap between normal and anomalous distributions. This makes F1AdaptiveThreshold unstable across runs. In addition, single-scale Perlin masks and alpha blending introduce boundary artifacts, and the current pipeline is disk-based with a required DTD download (~600MB).

I explored this through two prototypes. The first used multi-octave fBm masks with Poisson blending, which produced visually clean results but over-harmonized anomalies, making them too close to normal in feature space and reducing score separation. The second approach replaces this with fBm/ridged masks, soft-mask alpha blending, self-sourced donor patches with augmentations, and strength tiers to spread anomaly difficulty. This setup generates more diverse and better-structured anomalies without relying on external textures or disk writes.

The main impact is on score distribution. For the bottle category, the proposed method increases the score range from 0.0115 to 0.0413 (~3.6×) and the standard deviation from 0.0017 to 0.0072 (~4×), resulting in much better coverage around the decision boundary and more stable threshold estimation. This also showed better alignment with real test performance (validated with PaDiM).

Would appreciate your thoughts on whether this direction optimizing synthetic anomalies for calibration rather than just realism—fits Anomalib’s roadmap, and whether it makes more sense as a new generator or an extension of the current one. Happy to share code/notebooks if needed.

Thanks!
Abhay Kumar Das.
[LinkedIn]
[CV/Resume]