Finding parts others can't.

Anyone who's restored an older car has a story about The Part. The trim clip that's been discontinued for twenty years. The switch that only came on one trim level in one market. The bracket nobody has a part number for, because the only one you've ever seen is the broken one in your hand.

Finding that part is usually two problems stuck together. First, you have to figure out what it is. Then you have to find where it is. This Lab post covers the research behind Scavenja, which tackles both.

Problem one: what is this?

For common parts, identification is easy: look up the part number, or search your car's make and model plus a description. For the parts that matter most to restorers, it isn't:

• Part numbers are worn off, painted over, or were never stamped on at all.
• The same part was sold under different numbers by different suppliers, or superseded several times.
• Nobody agrees on what to call it. One person's "door handle escutcheon" is another's "that plastic ring thing".

A photo, on the other hand, is something everyone can produce.

Computer vision for parts

Recognizing car parts is a harder cousin of general image recognition. Many parts look almost identical across models, differing only in a mounting tab, a connector or a few millimeters of length. Photos are taken in garages, under poor light, of parts that are dirty, damaged or half-disassembled.

The approach we're researching combines several signals:

• Visual similarity: learning image representations where visually similar parts end up close together, so a photo can be matched against known parts even if that exact image has never been seen.
• Text in the image: reading any visible numbers, logos or markings, even partial ones, and using them to narrow the search.
• Context from the user: the vehicle, the area of the car and a rough description can each rule out large numbers of candidates.

The result is a ranked list of likely candidates with a confidence score, not a single answer. When it isn't sure, it should say so and ask for another angle or more context.

Problem two: where is it?

Once you know what you're looking for, the part could be almost anywhere: a major marketplace, a local classified ad, a specialist supplier, a junkyard's inventory, or a post in an enthusiast forum from someone clearing out their garage.

Each source describes parts differently. Listings are often poorly titled ("misc interior bits from my old project"), miscategorized, or missing part numbers entirely. Searching each one by hand, with several possible names for the same part, is exactly the tedious work that makes rare parts feel impossible to find.

Matching messy listings

The research here is about matching: deciding whether a listing is plausibly the part you need, even when it's badly described. That means normalizing part names and numbers, using cross-references and supersessions, and, where there are photos, comparing listing images against the identified part.

Then there's time. Rare parts appear rarely. The most useful thing a tool can do is keep looking after you've stopped, which is why saved searches and alerts are central to Scavenja rather than an afterthought.

Fitment is the real question

Finding a part that looks right isn't enough. What matters is whether it fits your car. We're exploring how to use OEM cross-references and fitment data across years, models and trims to flag compatibility, and to be explicit when the data isn't strong enough to be sure.

What we've learned so far

A few principles keep coming up:

1. Ranked candidates beat single answers. Restorers are experts. Show them the top options and let them decide.
2. Every clue counts. A partial part number plus the car's model year often narrows things down more than a perfect photo.
3. Persistence wins. For truly rare parts, a good alert is worth more than a great search.
4. Honesty about confidence builds trust. A tool that's occasionally unsure is more useful than one that's often confidently wrong.

We're continuing to work on recognition quality for older and rarer parts, and on covering more sources. If you've got a part that's stumped everyone you know, that's exactly the kind of challenge we want Scavenja to take on. Learn more about Scavenja.