Recod.ai/LUC - Scientific Image Forgery Detection | Baseline | Dinov2 training | Inference

In this video, I walk through my complete approach to the Recod.ai / LUC – Scientific Image Forgery Detection competition, where the goal is to detect and segment copy-move forgeries in biomedical images at pixel level.

This is not just another Kaggle-style competition.
The dataset is built from hundreds of confirmed forgeries from over 2,000 retracted scientific papers, making it one of the most realistic benchmarks in scientific image forensics.

I start by clearly explaining:

- The problem statement and why scientific image forgery is a serious issue
- The dataset structure and how the annotations are organized
- The evaluation metric (Optimal F1 + Hungarian matching + penalties) in a simple, intuitive way using a whiteboard

Then I deep-dive into:

- Why I chose DINOv2 as the backbone model
- The complete codebase structure and how the pipeline is organized
- Baseline implementation and inference flow
- Image pre-processing, sample forged images, and common patterns in manipulations
- Practical hints for improvement if you want to push performance further

This video is meant to give you both:

- Conceptual clarity (what problem we are solving and why)
- Implementation clarity (how to structure and build a real solution)

--------------------------------------------

🧠 What you’ll learn

- How copy-move forgery works in scientific images
- How pixel-level segmentation evaluation is done using optimal matching
- How to structure a research-grade ML codebase for competitions
- How to use foundation models like DINOv2 in real pipelines
- Where most baselines fail and how to improve them

--------------------------------------------

🕛 Timestamps

00:00 – Problem understanding
03:09 – Dataset understanding
05:03 – Evaluation metric understanding
07:03 – Why DINOv2 model
07:59 – Code base explanation
09:48 – Forged images samples
16:47 – Model code explanation
18:54 – Hints for improvements
20:41 – Baseline explanation
45:53 – Baseline Inference
51:50 – Image pre-processing results
54:48 – Conclusion message

--------------------------------------------

💻 Baseline Code Repository
https://github.com/HOTSONHONET/Recod.ai-LUC---Scientific-Image-Forgery-Detection/tree/master

--------------------------------------------

If you are interested in computer vision, medical imaging, scientific integrity, or deep learning competitions, this video will give you a solid end-to-end understanding of the problem and solution space.

If this helped you, consider liking, sharing, and subscribing — I regularly post deep-dive videos on ML, CV, and real-world AI problems.

Видео Recod.ai/LUC - Scientific Image Forgery Detection | Baseline | Dinov2 training | Inference канала AlgorithmsUntilRED