Cars Toyota Sera WIP

[norwichphoto]

Indian manufacturer Tata recently launched a sub-brand Tamo, the first offering is a bit Sera inspired...

https://www.tamo.co.in/racemo/geneva

Their car already in Forza Horzons 3

 

[Mr Whippy]

Here!

Thanks for that

The good (ish) old Max camera matching.

I've not had so much luck with it because you need to have points with known distances, and by definition ideally they're 'squared' off to each other.
So I got a very accurately shaped box and placed it in the car interior so I could camera match with it.

But then lens distortions caused me problems too, and often the box was too small within the shot (to the sides etc) so the main interior was in view

It gets you 'close enough' I suppose.

I've tinkered with using photogrammetry derived camera positions but still suffer with camera distortion issues. I need to do more testing because that method should be super solid... well worth using once I get it working

 

[Stereo]

I tried using a photogrammetry tool for Blender but it was way too strict about telling it where horizontal lines were in the photo (2 for X, 2 for Y to calculate the match and then I think it assumed they intersect at 0,0 so had to be in the same vertical plane) so it only really worked on photos of buildings.

Now I just do it on full manual, move/rotate the camera until it's close enough to help with the feature I'm working on. Generally I start by putting the 3d cursor on a point I can see in the photo, then rotating/scaling around that point and if it doesn't fit I tweak the fov and try again. I found a GIMP plugin that can undo barrel/etc. distortion of known cameras so I also have the option of using that if a photo has full EXIF data, likewise if the EXIF data has accurate focal length I can just set it (I found that for example iphones only list this as a whole number and their sensor size is so small it's useless, eg. you'll get 4mm focal length 6mm sensor, meaning it's actually somewhere 3.5-4.5 which is a huge range of fovs)
Some internet photos work fine, some just have no chance of matching up, I imagine it's due to various cropping/scaling choices that don't make it into EXIF. I find it's not really worth trying to compensate for that sort of thing, way too many degrees of freedom. If a photo's not working just move on.

 

[A3DR]

Using a box to do the matching is always a good idea (by box I mean a very basic shape of the object you're modeling), could be frustrating at first but after a while you get a better idea where to place the points and it becomes an easier process.
I usually do it at the beginning of a project, once I get a close enough basic shape using blueprints or pictures, then before doing the detailing I check again with other views.

It's not always 100% accurate, but helps you to get the correct proportions of the object in real life.

Here's an example of a detail, I've adjusted the camera by hand just to get a reference for the shape inside the tail light. It doesn't perfectly match but it's good enough to get the depth right:

That's how it looks with reflections:

 

[Mr Whippy]

Ah of course, built it off BPs quite roughly, then you can use photomatch points dimensionally by picking obvious features that are present in photos and bps.

That makes sense to me.

I'm having a play right now using photogrammetry (photoscan) to undistort photos and get cam locations, but then use them for photo matching data.

I'd tried this previously but failed to note how significant my focus breathing effect was (so impossible to fully compensate for varying distortion via photogrammetry), but have just done a quick test earlier to check with a static focus too!
If things look good I'll post a how-to.

As your results prove A3DR, even in a world of photogrammetry, laser scanners, and CAD models, photo modelling and cam matching are still top end tools and in some cases all we have access to.

Having some toolsets or workflows that make cam matching and undistorting really easy would no doubt help anyone making a car (especially ones where *ALL* you have is photos off the internet, and no BPs!)

Dave

 

[Mr Whippy]

I tried using a photogrammetry tool for Blender but it was way too strict about telling it where horizontal lines were in the photo (2 for X, 2 for Y to calculate the match and then I think it assumed they intersect at 0,0 so had to be in the same vertical plane) so it only really worked on photos of buildings.

Now I just do it on full manual, move/rotate the camera until it's close enough to help with the feature I'm working on. Generally I start by putting the 3d cursor on a point I can see in the photo, then rotating/scaling around that point and if it doesn't fit I tweak the fov and try again. I found a GIMP plugin that can undo barrel/etc. distortion of known cameras so I also have the option of using that if a photo has full EXIF data, likewise if the EXIF data has accurate focal length I can just set it (I found that for example iphones only list this as a whole number and their sensor size is so small it's useless, eg. you'll get 4mm focal length 6mm sensor, meaning it's actually somewhere 3.5-4.5 which is a huge range of fovs)
Some internet photos work fine, some just have no chance of matching up, I imagine it's due to various cropping/scaling choices that don't make it into EXIF. I find it's not really worth trying to compensate for that sort of thing, way too many degrees of freedom. If a photo's not working just move on.

Good thoughts there. Crops not showing up in EXIF is indeed a big problem.

It might be worth starting a thread with different approaches to cam matching, cameras, reference images etc?
I'd hate to spam this thread too much more with camera matching brain dumping

 

[Stereo]

Using a box to do the matching is always a good idea (by box I mean a very basic shape of the object you're modeling), could be frustrating at first but after a while you get a better idea where to place the points and it becomes an easier process.

Ah right I forgot to mention those steps.
First I do a literal box the size of the car (length/height/width no mirrors, wireframe only) and a box the size of the wheelbase (track width+length, height equal to half the tire so its top edges are the axles) and set up whatever blueprints I can. Usually blueprints need to be smaller than the box because they don't draw the shallow curvature of front/top/back correctly so matching wheelbase is key. Low-fov side/front photos can work with the same limitation, won't be accurate around the edges. If a car's been raced these are often easy to find since people photograph them with long telephotos.

If no blueprints then I align initial photos using the wheels, block in the hood or something, and fix the alignment. Mirrored features are your friend in this part, once you have the FOV on a photo correct you can get something positioned in 3d by having it match while mirrored.

 

[A3DR]

Ah of course, built it off BPs quite roughly, then you can use photomatch points dimensionally by picking obvious features that are present in photos and bps.

That makes sense to me.

I'm having a play right now using photogrammetry (photoscan) to undistort photos and get cam locations, but then use them for photo matching data.

I'd tried this previously but failed to note how significant my focus breathing effect was (so impossible to fully compensate for varying distortion via photogrammetry), but have just done a quick test earlier to check with a static focus too!
If things look good I'll post a how-to.

As your results prove A3DR, even in a world of photogrammetry, laser scanners, and CAD models, photo modelling and cam matching are still top end tools and in some cases all we have access to.

Having some toolsets or workflows that make cam matching and undistorting really easy would no doubt help anyone making a car (especially ones where *ALL* you have is photos off the internet, and no BPs!)

Dave

It all comes to the basics as you can see. I'm aware there are new tools out there but why make a mess when things can be simple with good results. In my case, this kind of work flow works out quite well, there's also an important factor here and it is that we're not aiming for 100% accuracy, I think my latest models must be somewhere around 90-95% compared to the real car, but only the most trained eyes can spot those differences.
There's also the fact that we need to know where to draw a line, otherwise we could spend countless hours working on the same thing and never get it finished.

Ah right I forgot to mention those steps.
First I do a literal box the size of the car (length/height/width no mirrors, wireframe only) and a box the size of the wheelbase (track width+length, height equal to half the tire so its top edges are the axles) and set up whatever blueprints I can. Usually blueprints need to be smaller than the box because they don't draw the shallow curvature of front/top/back correctly so matching wheelbase is key. Low-fov side/front photos can work with the same limitation, won't be accurate around the edges. If a car's been raced these are often easy to find since people photograph them with long telephotos.

If no blueprints then I align initial photos using the wheels, block in the hood or something, and fix the alignment. Mirrored features are your friend in this part, once you have the FOV on a photo correct you can get something positioned in 3d by having it match while mirrored.

That's what I do as well, most of the time I build the BP's in photoshop using a combination of pictures and drawings (if any available). I compensate the edges issue with camera matching, that's why I start with a rough shape so it's easily manipulable with FFDboxes mostly.

 

[Mr Whippy]

A3DR, I completely agree.

I also use the same technique Stereo has outlined, roughing out from there.

But like you noted, you can end up going around in circles trying to use data explicitly, rather than implicitly.

Aiming for 100% is probably 10x the investment than getting 90-95% as you say.

Finding fault at that level is probably near impossible without overlapping CAD data vs your model, or some really observant real life owner looking for faults!

I just like the explicit exercise. It can take some of the fun out of a workflow, but retopo off CAD can just 'get the job done' which is sometimes what you want.
Having CAD data is a luxury, the angle I'm coming from is to try find more explicit ways of working right down the pipeline... back to this classic workflow of photo-matching.


Either way your work is stunning and more than good enough for me. And thanks for the inspiration because it truly is inspirational and is making me spend more time working on my projects!

 

[A3DR]

Unexpected.

Still WIP

 

[ScuzzySera]

Needs the roof spoiler Fantastic work!

 

[After_Midnight]

Unexpected.
View attachment 184190
Still WIP

Unbelievable

 

[DutchDevil79]

Wow pimp my ride extreme. haha

 

[A3DR]

Yeah I almost add a roof spoiler! but I've said to myself this was not going to be something that'll take more than a day, so here it is:

Full gallery:
https://a3dr.artstation.com/projects/1YRzG

Not gonna go full ape on the interior, sponsors are BS, body kit doesn't exists (designed myself) and this looks ridiculous, but that's the whole point it's going to have a tuned MR2 engine so it's going to be fun to drive, that's for sure.

 

[Joseph Maxx]

That thing looks like an absolute beast, I love it! Too bad about the body kit not existing, one of the more interesting ones I've seen, maybe you can start a side business

 

[Hotdogxx60]

Maybe a slim line bonnet scoop or intake for better cooling otherwise great the way it is.

 

[GTAce]

Holy crap, hope it sounds as brutal as it looks. Awesome work man.

 

[norwichphoto]

love the time attack Sera. My personal collection has doubled to two cars. Just bought Assetto Corsa as it was half price on Steam.

 

[A3DR]

I'm still waiting on the final physics from Joseph, so you'll be able to drive all three versions soon, hopefully!

 

[norwichphoto]

If I had the time & money I'd do the Time Attack version for real. However, both my cars are getting restored to standard specification. I'm keeping one and selling the other.