Ongoing tests of candidates for the Infragram Webcam have gone relatively poorly to date, but over the weekend I received a number of samples from I Zone Technologies and while I was initially a bit intimidated at the prospect of converting them, I Zone staff helpfully recommended that the best way to remove the filters was to unscrew the lenses, even though they were inside auto-focus mechanisms (see above). The models I dealt with were the UMBMU and the UM2CU, 5 and 2 megapixels, respectively. They are TINY.

I used a VERY fine pair of tweezers with an angle, held the indentations in the plastic surrounding the lens, and rotated the circuit board around it, holding the lens still.

Amazinglly this worked for both cameras, and I didn't have to resort to melting the glue and removing the whole autofocus assembly. Instead the lenses came out without too much trouble, although I occasionally squeezed the tweezers too much and sent it flying across the desk.

The harder part was removing the filter, which was recessed on both models. I ended up cracking it carefully with a knife and picking out the parts, blowing it clean. They won't have to do this at the factory :-)

The converted camera with an external infrablue (Rosco #2007) filter:

Here's the higher-resolution UMBMU:

Imaging test results

Whoops! I almost forgot to post the results of the images. I really need a better test setup but the UMBMU seemed to do quite well (click to see full results):

And I think we can improve these, but here are the UM2CU 2 megapixel results:

That last one clearly has poor white balance, and is detecting a lot of blue light. I'm going to try messing with it in Infragram Sandbox to see if I can draw out some better results.

Long story short, not 100% but MUCH better than any of the previous cameras I'd tested. I now feel much more confident about the webcam edition!

Update: Infragram Sandbox results

Actually this worked pretty well; I used the Infragrammar equation (R-B/2)/(R+B/2)*3 which just halved the blue channel and generated NDVI, then tripled the results. It's a bit garbled but does seem to clearly differentiate vegetation from the background. This may necessitate more testing but seems promising.