AI Visibility for Instant Noodles: Fatal Display Scoring Mistakes

25/10/2026

There is an interesting paradox in the Vietnamese FMCG industry: we are living in the era of AI Vision, where computers can read license plates or count foot traffic with high accuracy, yet when it comes to counting instant noodles on supermarket shelves, many businesses remain 'blind' to their actual market coverage. Why? Because the input data is too dirty and the expected output is too vague. For an instant noodle brand, every day spent without knowing your shelf position is a day of losing market share to competitors. The problem is not the technology, but how we frame questions for that technology.

Mistake 1: Using a generic recognition model for all packaging

Many assume that if AI can recognize a human face, it can recognize a noodle pack. Wrong. Instant noodle packaging in Vietnam is complex: from 3,000 VND to 25,000 VND packs, with similar colors and faint printed images due to store lighting. Using a generic AI Visibility model causes recognition error rates to spike, especially when confusing product lines from the same brand. The result is a beautiful report with useless data. The fix is to train separate models for each packaging type, and even for each geographic region, as lighting and photo quality vary by location.

Mistake 2: Ignoring shooting angles and distances

Technicians photographing shelves often stand too far away or hold their phones at an angle. Display scoring systems rely on the principle that objects must occupy a sufficient portion of the frame. If a noodle pack only takes up 2% of the image area, accuracy drops significantly. I have seen a project report 85% accuracy simply because 30% of the photos were blurry or taken at an oblique angle. Do not expect AI to see what the human eye cannot. The data collection process must be standardized: distance, angle, and lighting.

Mistake 3: Expecting absolute accuracy from field data

This is a psychological error. We want AI to deliver 100% accuracy. But in reality, VTraks reports 99.97% accuracy according to vtraks.com, an impressive figure but still a statistical one. In practical deployment, out of 1,000 images, there will still be 1-2 controversial ones. The error is not in the AI, but in the lack of a review process. A good AI Vision system needs a feedback loop: when the system is uncertain, it must flag the image for human confirmation, thereby improving the model for the next time. Do not let the system make decisions when it is in doubt.

Mistake 4: Lacking competitive context

Display scoring is not just about whether your product is present. It is about where your product is relative to competitors. If your noodle pack is pushed to the bottom shelf while competitors occupy eye-level space, your display score is low, even if the shelf is full. Many systems only count quantities, not positions. This leads businesses to believe they are controlling the market, when in reality they are being 'squeezed' for display space. You need an evaluation algorithm based on visibility and prominence.

Mistake 5: Not validating image integrity

In a multi-tier distribution environment, from large supermarkets to small grocery stores, salespeople may take screenshots, use old photos, or even submit forged images to bypass monitoring. Without mechanisms to detect duplicate or screenshot images, you are building on sand. Input data must be cleaned before processing. This is a step many systems skip, assuming that 'if it is a real photo, no check is needed.' Remember, humans have motives to cheat, and AI cannot judge those motives unless it is designed to detect anomalies in the pixels.

Mistake 6: Inconsistent infrastructure deployment

You may have a good recognition system, but the processing speed is too slow. If each image takes 10 seconds to analyze, the inspection process becomes cumbersome, and employees will find ways to bypass it. Modern systems need to process images in under 1 second to create a seamless experience. Additionally, integrating data into a centralized dashboard is crucial. If data is scattered across different Excel files, you cannot see the big picture. A platform like AIVISION, with products like VTraks, helps consolidate this data, allowing managers to see trends rather than just isolated numbers.

What remains unsolved?

The technology is ready, but people and processes are not. We still lack a national standard for display data quality. We need more open datasets to train better models, especially for emerging markets like Thailand or the Philippines, where packaging and shopping habits differ. We also need more tools to train sales teams to understand why AI scores things the way it does, rather than just blindly complying. AI is a support tool, not a judge. We need to continue building trust in the system through transparency and the explainability of machine decisions.

There is more here than one article can hold. Keep reading on the AIVISION blog, look at our display scoring solution, or get in touch.