Sidewalk Gallery: label placement different on small card vs. expanded modal

[jonfroehlich]

You'll notice that the blue label location is different on the small card (drawn on street) vs. on the expanded card:

[jonfroehlich]

Looks like this is pretty common now:

I had not noticed this before. I wonder if a recent change caused the problem?

[jonfroehlich]

But it's not always the case:

[misaugstad]

@jonfroehlich I was able to reproduce this going back to our earliest version!

[misaugstad]

Oh right, this is really just this issue that I've been needing to fix 😅 #2485 It's just the issue of not making the crops correctly either. Projecting onto the static SV image is just not working correctly yet.

[jonfroehlich]

Gotcha. I'm gonna go ahead and make the video then. And we'll deal with this when we can. Sound good?

[misaugstad]

Perfect, yep!

[michaelduan8]

Yeah, definitely might be worth doing a deeper analysis (expand on our regression analysis) to gain more accuracy in both crops and label placement on static images.

[misaugstad]

might be worth doing a deeper analysis (expand on our regression analysis) to gain more accuracy in both crops and label placement on static images

I think the regression analysis you're talking about was in estimating the lat/lng of labels, not their location on the images and the crops. The bad crops seem to more likely be a problem with our math that desperately needs to be resolved :)

[jonfroehlich]

Here's a video example. I'm hoping @uditpatwal's work will help solve this.

LabelJumpingAround.mp4

[misaugstad]

This is a sub-problem of #3095.

The reason that this is happening is because GSV's static API, which we use to get images for the small cards, has a maximum image size of 640 x 480 pixels while we are actually requesting images with dimensions of 720 x 480 (bc those are the dimensions of the Explore page GSV window).

In #3173 I made some small improvements to make the location of the labels on the cards much more accurate, but there is still going to be slight discrepancies until we deal with #3095.