I've been spending the last few months mapping a problem space. Anti-counterfeiting. Specifically, whether computer vision is the right fit for automating product authentication at scale in logistics and retail.
The short answer is yes, with conditions.
The problem
Counterfeiting is a large and underserved problem in physical goods. Luxury brands, consumer electronics, pharmaceuticals, sportswear. The estimates of market size vary and most of them are unreliable, but the directional claim that it's significant is not contested. The brands losing revenue to it are serious about solving it.
Current solutions are mostly manual: inspectors who know what authentic products look like, physical authentication tags that can themselves be counterfeited, and spot-checking processes that scale poorly. The rate of false negatives (counterfeits that pass inspection) is high because humans get tired, lighting conditions vary, and the volume is large.
This is a CV problem. The conditions that make it hard for human inspectors are conditions where automated visual inspection should perform well: high volume, consistent category of objects, need for consistent application of a standard. The failure modes are different. A trained model doesn't get tired at hour six. It does fail on distribution shift.
Why YOLO specifically
The inspection context matters. In a logistics environment, you're not doing controlled lab photography. Products come through on conveyor belts, in variable lighting, partially obscured, at different angles. The inspection has to happen fast enough not to create a bottleneck.
Real-time detection is the constraint. YOLO (You Only Look Once) trades some accuracy for speed by doing detection in a single pass through the network rather than the two-stage approach of something like Faster R-CNN. For a logistics use case where throughput is the constraint and you can design the hardware setup to reduce variability, the speed advantage is worth the accuracy tradeoff.
The alternative is a classification-only approach where you assume the object is centered and correctly framed. That requires controlled input conditions that logistics environments can't reliably provide. Detection first, then classification, is the more robust architecture.
What I don't know yet
The accuracy requirements are unclear. What rate of false positives is acceptable? A false positive (flagging an authentic product as counterfeit) has real costs: slowing logistics, damaging customer relationships. A false negative (passing a counterfeit) has different costs depending on the category. The right accuracy target is a business question, not a technical one, and I haven't answered it yet because I haven't talked to enough people in the domain.
I've done about twenty conversations so far with people in brand protection, logistics operations, customs, and a few brand IP teams. The picture is consistent on the problem and inconsistent on the solution requirements. I need more conversations before I write a line of code.
What I'm taking from the Zillion Pitches experience
At Zillion Pitches we built the technical infrastructure before fully understanding whether our customers would pay for the output. The unit economics reflected that. This time I'm doing it the other way: validate the problem and the willingness to pay first, then build.
The CV work is genuinely interesting to me. I've been reading the YOLO papers, running experiments on public datasets, mapping the architecture. That work is useful as preparation and premature as execution. The conversations come first.
With gusto, Fatih.