Quality Match GmbH

🔍 Struggling with accurate quality assurance? The traditional "Full Image QA" approach has experts look through every annotation, identifying errors like incorrect labels, missing objects, geometric accuracy and more. It's time-consuming and error-prone. 👁️🗨️ Try checking all relevant aspects in the left image within 1 minute. Difficult, right? Now have a look at the right image. 💡 The solution is as simple as the solution itself: using one-dimensional properties for each single Quality Aspect. For example, one Quality Aspect of a pedestrian detection label specification could be ensuring each detected object contains a living human being. Another could be that the bounding box surrounding that object is tight, touching the object at all four sides. ✅ A good and EU AI Act-compliant quality assurance process, as per ISO 9001 and the upcoming ISO 5259-2, needs to be reproducible and auditable. That's what we do at Quality Match.

Please reach out to me in case you want to learn more about how you can bullet-proof your computer vision application - we can supply you with both training data as well as model QA tools.

Don't forget about the EU AI Act! Reach out to Quality Match GmbH to see whether your safety critical computer vision application needs a checkup! https://lnkd.in/ek5ccXh6

Two weeks have passed since our major AI release and the work on making our AI Nano-Tasks even better has long started - thanks to the valuable feedback of our test users! 🤝 You'd also like to find out if you can benefit from our AI feature? 👉 Get in touch and we make sure that you get your free test account 👈 ❓Did you know that with our AI Nano-Tasks you not only can automatically investigate any quality aspect like a full team of human annotators would, but you also can enrich your existing annotations with attributes like occlusion levels? 🤯 And thanks our repeat system the quality assurance for this enrichment is built in!

ℹ Why you should re-think your annotation QA process In traditional Quality Assurance (QA), the adder-approver method — where an adder annotates a frame and an approver reviews it — has been the standard for years. This process often results in errors due to the complexity and volume of decisions each frame demands. Annotators must take up to 400 decisions per frame, keep extensive labeling guides in mind, and work under tight time constraints due to budget limits. At Quality Match GmbH, we are transforming this model with our innovative approach utilizing AI-driven Nano-Tasks. By decomposing complex questions into simpler, manageable tasks, we enhance the entire QA process in several key ways: ➡ Focus and Simplicity: Each nano task concentrates on a single decision, drastically reducing complexity and minimizing errors. ➡ Auditability: Every annotator’s response is documented and the QA process is auditable. ➡ Statistical Analyses: By repeatedly annotating individual nano tasks, we can conduct detailed statistical analyses. ➡ Resolution of Ambiguities: Nano tasks help pinpoint and resolve ambiguities and edge cases more effectively. Have a look how simple QA can be:

Wednesday we talked about data quality and issues in annotations for four common quality metrics. 1️⃣ False Negatives or objects that have been overlooked 2️⃣ False Positives or hallucinated objects 3️⃣ Attribute correctness 4️⃣ Geometry correctness 🔎 Did you find at least 5 errors in the pedestrian detection example? Swipe ➡ to see some results! Find the original post with more details here https://lnkd.in/eAbK2rtJ

🗣 People often talk about "data quality" but what does this actually mean? ▶ At Quality Match GmbH, we are experts at choosing the right metrics and helping you to find an ideal process that balances cost, speed and quality just like you need it in your current computer vision project. We can apply these metrics either to human-labelled data or to the results your product is already producing. ◀ So, what do we recommend? 👇 In computer vision, we can find at least four types of quality metrics: 1️⃣ False Negatives or objects that have been overlooked - sometimes objects are simply ignored by models and annotators alike because they are too far away, too blurry or too heavily occluded. But did you know that a common problem is also that the object to be found is too large? 2️⃣ False Positives or hallucinated objects - AI models make this mistake because they are often trained to be more sensitive to detections so they can be filtered out further downstream in the vision product. Humans often make these mistakes by either misunderstanding the taxonomy (a police person should not be labelled as pedestrian, it is an official person instead) or by simply accidentally clicking on the wrong class in an user interface. 3️⃣ Attribute correctness - highly related but often overlooked are attributes such as occlusion amount or orientation of an object. Verification of these attributes is very often forgotten when you just show an image with 20 bounding boxes to a quality assurance person and ask: "Are all objects correctly labelled according to the label guide?" 4️⃣ Geometry correctness - this can be on polygons, segmentations, bounding boxes, and all other geometries. Here the main challenge is that the rules for "correct geometry" are often very vague and highly subjective. What do you do with an occluded object that is also heavily blurred? Can you clearly outline the object in a fully unambiguous way? Here it is very important to have a clear and objective approach that captures the ambiguity rather than assuming that only one single truth exists. ❓Now that you know everything about data quality metrics, here is a challenge for you: find at least 5 errors in this pedestrian detection training data:

The past weeks have been busy between the #CVPR2024 in Seattle, a workation in Mallorca and the biggest product release of our company - so far. ⛱👩💻 Our #workation in Mallorca was centered around team-building and intense testing of our newly released AI feature, complemented by great food and a swimming pool. 🚗 🏥 🚜 The CVPR was a huge success, presenting our paper and poster about data ambiguity and enjoying technical discussions on how to improve quality assurance processes in #ComputerVision applications across industries. 🤖 And then the release of our #AI Nano-Task Alpha: you now can use our platform to automate your QA and attribute enrichment processes, while getting all necessary insights in your data that you need to take further actions. Test it for free! Reach out if you want to learn more about how we do things at Quality Match GmbH.

🚨 AI Product Release 🚨 We are very excited to announce our latest feature within our annotation QA platform HARI: the AI Nano-Tasks, version Alpha 🎉 Our teams worked relentless during the past months to make this possible and now you can experience it free of charge to test it! Some Benefits of AI-Nano-Tasks Alpha 📉 Reducing Quality Assurance Costs: Improve your annotations through automated QA and enrichment based on human decision-making, while reducing the required QA budget significantly. E.g. by 27% (100k Annotations) or 81% (1M Annotations). 💪 Leverage your Foundational Models: Iteratively improve your model’s performance through targeted automated QA. 🔎 Enhanced Statistical Insights: train your AI on input from several human annotators. Instead of your conventional 2-3 layered QA your AI Nano-Task performs as 10 or more layers, at a fraction of the conventional cost. ❓ Ambiguous Data Handling: Automatically identify and manage difficult annotations with conflicting answers, deciding on further human checks or data removal as we have shown in our paper at the CVPR 2024. ➕ Enriched Annotations: Use AI Nano-Tasks not just for QA but also to add attributes with built-in quality assurance. And since this is an Alpha version you can influence the future development: Your feedback is highly valuable to us in defining our future direction!

🚨 AI Product Release 🚨 We are very excited to announce our latest feature within our annotation QA platform HARI: the AI Nano-Tasks, version Alpha 🎉 Our teams worked relentless during the past months to make this possible and now you can experience it free of charge to test it! Some Benefits of AI-Nano-Tasks Alpha 📉 Reducing Quality Assurance Costs: Improve your annotations through automated QA and enrichment based on human decision-making, while reducing the required QA budget significantly. E.g. by 27% (100k Annotations) or 81% (1M Annotations). 💪 Leverage your Foundational Models: Iteratively improve your model’s performance through targeted automated QA. 🔎 Enhanced Statistical Insights: train your AI on input from several human annotators. Instead of your conventional 2-3 layered QA your AI Nano-Task performs as 10 or more layers, at a fraction of the conventional cost. ❓ Ambiguous Data Handling: Automatically identify and manage difficult annotations with conflicting answers, deciding on further human checks or data removal as we have shown in our paper at the CVPR 2024. ➕ Enriched Annotations: Use AI Nano-Tasks not just for QA but also to add attributes with built-in quality assurance. And since this is an Alpha version you can influence the future development: Your feedback is highly valuable to us in defining our future direction!