2026-09-11 –, Main stage
Your model isn't the problem. Your data is.
In production computer vision, teams consistently reach for model-level fixes, bigger backbones, longer training runs, more data, while the underlying dataset stays messy, mislabelled, and unrepresentative of the real world. We did the same, until we stopped.
This talk presents three data strategies we applied to a single, deceptively simple problem: counting crates moving through a logistics environment. The problem looked straightforward. The data was not.
The three strategies are:
- Annotation quality at scale: using model-assisted labelling and embedding-based clustering to systematically surface and correct errors hiding in production datasets
- Synthetic data generation: covering rare but critical edge cases absent from real captures, and navigating the domain gap between synthetic and real images
- Intelligent dataset reduction: selecting maximally informative training subsets, based on the counterintuitive principle that more data is not always better data
This is not a benchmarking talk. There are no leaderboard numbers here. The goal is to share a way of thinking about data that you can apply to your own problems, and to be honest about where each strategy works, where it struggles, and how they interact.
Target audience: Data scientists, ML engineers, and applied researchers working with real-world computer vision or ML pipelines.
Takeaways: A practical, transferable framework for approaching data quality.
Overview
This talk argues that production computer vision performance is fundamentally a data problem, not a model problem, and presents three data-side strategies to back that claim. Rather than showcasing results or benchmarks, we share the principles and workflows we believe are essential for any practitioner deploying computer vision in the real world.
The entire talk is grounded in a single, concrete case: counting crates moving through a logistics environment. We chose this case precisely because it sounds trivial. It is not. The gap between "this should work" and "this actually works in production" turned out to be almost entirely a data gap, not a model gap. To increase performance in production, we focused on improving our dataset instead of augmentations, backbones, or hyperparameters.
The three strategies are presented as complementary principles, not isolated techniques. Each one is explained through the lens of this case, with honest discussion of what worked, what didn't, and what we'd approach differently. The talk is designed so that attendees walk away with transferable thinking they can apply to their own pipelines, regardless of domain or framework.
Outline with Estimated Timing (45 min)
Opening & framing (5 min)
We open with the case itself: crates on transport dollies moving through a logistics facility. The task sounds simple, count what you see. We use this apparent simplicity as the hook: what happens when you deploy this in production, where crates are stacked unpredictably, mixed with foreign objects, and seen under conditions your training data never captured? This sets up the central thesis: the model was never the bottleneck. The data was. We frame the three strategies as essential workflows for production-grade computer vision.
Strategy 1: Annotation quality at scale (12 min)
Production datasets are systematically mislabelled, and teams rarely acknowledge it. We show how to use the model's own predictions to surface annotation errors using confidence disagreements and embedding-based clustering to find the most ambiguous, error-prone samples. We walk through the iterative review workflow: when to inspect, what to correct, and when to stop. The key principle: annotation quality is not a one-time task but a continuous loop that should be integrated into your production pipeline.
Strategy 2: Synthetic data generation (10 min)
Some edge cases simply don't exist in your real dataset: new crate types, unusual stacking configurations, lighting conditions you haven't encountered yet. We explain how synthetic data can fill these gaps, the decisions involved in scene design and rendering, and the challenge of bridging the domain gap between synthetic and real images. We're candid about the limitations: synthetic data doesn't transfer automatically. We also address one honest exception to our "data first, model second" thesis: to close the domain gap between synthetic and real images, we use a domain adversarial training approach that is, strictly speaking, a model-side intervention.
Strategy 3: Intelligent dataset reduction (10 min)
More data is not always better data. In production, you are continuously collecting vast amounts of unlabelled data. Annotating all of it is both impractical and counterproductive: most samples are highly similar and add little to the model's performance. We present the principle of curating maximally informative subsets from large uncurated pools, selecting what to train on rather than training on everything. We show results from our own pipeline where a fraction of the dataset matched or outperformed the full set, and discuss the trade-offs of aggressive curation and how to manage them.
Lessons learned & closing (8 min)
We reflect on what was harder than expected, how the three strategies interact (and sometimes conflict), and what we'd do differently. The closing message: these strategies form a continuous data pipeline, not a one-time fix. Production is never "done," but these workflows make your life significantly easier when you're running a computer vision solution in the real world. Open floor for questions.
Additional Details
Prior knowledge expected: Familiarity with supervised learning and basic object detection concepts (bounding boxes, confidence scores). No specific framework or tooling knowledge required.
Talk type: Principles-based, case-study grounded. No live demos or benchmark comparisons. The goal is conceptual clarity and transferable judgment, not replicating a specific stack.
Scope boundaries: This talk focuses exclusively on data-side interventions. We do not cover model architecture choices, deployment infrastructure, or real-time inference optimisation.
Materials: Slides will be shared publicly after the conference.
Alexander Kern is AI Solutions Expert at Clockworks where he works on numerous computer vision solutions that derive structured insights from raw pixel data. His work exemplifies how AI should be combined with smart engineering to deliver solutions that always and reliably work in production.
I’m an AI Solutions Engineer at Clockworks in Rotterdam. I studied AI and Data Science in Nijmegen, where I developed a passion for computer vision. Right after graduating, I joined Clockworks, where I’ve been building AI systems for real-world applications for the past four years.
Among other things, I’ve contributed to the Vision AI technology behind the Stack&Track solutions and our own product, Blicker. Through this work, I’ve seen how challenging it is to build systems that perform reliably outside controlled environments.
I’m especially interested in improving data quality, building better annotation workflows, and developing practical Vision AI systems that hold up in production.