Commons:Featured picture candidates/File:Line scan photo of nine car BART C1 train in 2017.jpg/3

From Wikimedia Commons, the free media repository
Jump to navigation Jump to search

File:Line scan photo of nine car BART C1 train in 2017.jpg, featured[edit]

Voting period is over. Please don't add any new votes.Voting period ends on 16 Jun 2018 at 02:35:01 (UTC)
Visit the nomination page to add or modify image notes.

A BART train in San Francisco


Scrollable version of above
  •  Comment I am aware the colours aren't perfect currently, but they are vastly better than before. I've spent a lot of effort trying to mitigate the tinge and I think that now they have been reduced to a point that doesn't significantly detract from the image. As for holding up the colorchecker passport, remember the line scan camera can only see a tiny 1px slice of it. There is no way to image the whole passport at once. Moreover, every pixel has a slightly different sensitivity. For example, the greenness is much stronger towards the top of the image than the bottom (as seen between cars 349 and 1644). I think you are probably right that calibration in the field would be better than calibrating at home since the sensitivity and noise level of the sensor seems to change with temperature. If you can think of a way to move the colorchecker passport in such a way that it is possible to programmatically recover which of the squares each pixel is looking at, such that each pixel has the chance to look at all the squares, all the while not changing the angle of the colorchecker passport with respect to the light source (assuming the colorchecker passport isn't perfectly Lambertian), while moving it fast enough that the thing can be captured without a significant change in the sun's position --- please suggest such a method so that I may implement it. By the way, of course the colours are less accurate for dark areas. Like all image sensors, this is a (nearly) linear image sensor, whereas the image shown here is a gamma-corrected image using the piecewise sRGB function. The slope of the sRGB function, as you may recall, is 12.92 in the dark region. Any constant additive noise in the sensor, such as thermal noise, in that region will be amplified by 12.92 times. If you consider the fact that the shutter speed used here (around 1/30,000 s -- I'll have to check when I get home) is several times faster than the fastest possible shutter speed of your DSLR camera, you'll see that the noise level is actually fairly reasonable. dllu (t,c) 07:17, 8 June 2018 (UTC)[reply]
  •  Info Here is a link to the 4096 calibration matrices (3x3) that I'm using. If you find more accurate calibration matrices for this particular photo, feel free to make a pull request. dllu (t,c) 07:21, 8 June 2018 (UTC)[reply]
  • How about a compromise? I guess your hands are free while ur device scanlines the train. You could take pictures of the passing train with a "regular" camera (which would be easier to calibrate using the color checker) and use the outputs as reference (don't ask me the details, I don't know how). Sounds feasible for someone versed in image processing. - Benh (talk) 17:41, 8 June 2018 (UTC)[reply]
  • My colour checker idea was a bit of a joke really. But Benh's idea is a good one. For most purposes even, getting the colours similar by eye would be good enough -- a whole lot better than trying to remember the sky colour or shade of a poster blue. Wrt variation between pixels, surely that is something you can calibrate at home in the studio with reference images -- isn't the variation in the field a global effect? Wrt your shutter speed and amount of light captured, do you have any idea what the equivalent ISO of a regular camera would need to be? Does your camera need focused or is it like a pinhole camera? -- Colin (talk) 08:32, 9 June 2018 (UTC)[reply]
  • The variation seems to change with temperature and time or something. Now that I think about it, the green cast can also be due to lens flare since the sun was in the background. I calibrated the camera at home but it was several months after the photo was taken, and the greenish cast towards the top of the image didn't seem to show up in my calibration data. And yes, the camera needs focusing. It just uses a regular photographic lens (Nikon F mount, manual focus). Nailing the focus of a large aperture lens when you only have a tiny sliver to look at is incredibly challenging (I think this was shot at f/2.8). The length of the sensor is 28 mm, about the same as the diagonal length of APS-C sensors. I spent all afternoon scanning BART trains and this was the best one I got. By the way, in case you were interested: here's a video of me scanning a Queensland SMU 260 recently. I'll process and upload those soon. Uunprocessed, uncalibrated sneak peeks SM260 SM220. dllu (t,c) 09:07, 9 June 2018 (UTC)[reply]
Confirmed results:
Result: 15 support, 0 oppose, 0 neutral → featured. /--Ikan Kekek (talk) 09:31, 12 June 2018 (UTC)[reply]
This image will be added to the FP gallery: Objects/Vehicles/Land_vehicles#Rail_vehicles
Retrieved from "https://commons.wikimedia.org/w/index.php?title=Commons:Featured_picture_candidates/File:Line_scan_photo_of_nine_car_BART_C1_train_in_2017.jpg/3&oldid=305991178"