Perhaps a mix of tiles, bigger objects and full backgrounds?

Perhaps it's that simple? I'm sure it would take a lot of fine tuning to get it too look just right... but I suppose that's the kind of work it takes to make a high quality commercial game.

Have you played the game though? Because there's a lot of places that simple don't seem tiled...

It's just that there are so many places in the game that don't seem to line up like tiles. Either that or they have a lot of transition tiles.

Its pretty common for maps to have their own backgrounds that are just big bitmaps. There isn't a real benefit to tilling those.

As for the slopes, you'll notice in those screenshots they're all 45 degrees. The slopped tiles are just special tiles. Nothing too advanced about 45 degree tiles, even Mario had that.

For a 45 degree slope you can do two things. One approach is to add two flags for each tile, one meaning a slope going up and one meaning a slope going down. When you do collision detection you just check the flag and if its set you instead check for collisions against the tile as if the upper left half of it wasn't there.

Another approach is to have all your tiles be triangles. You get a square tile by combining two triangles. Depending on the amount of extra data you associate with tiles, this method can be less memory efficient then the first.

Once you've discovered you've collided with a slope you simply decrement the y value of the new position until the new position isn't colliding anymore.

Got a screenshot? I'm curious to see it.

I'd say from what I recall is that they use tilemaps for all the foreground items and huge 3D like bitmaps or even possibly 3D for the background which give it a great feeling of depth and immersion. If you've ever played strider2 (the arcade version) or megamanX 7 you'd see what I mean as they do a similar approach.

For old school platform games you don't need realistic physics anyway. However.. you can calculate an approximate normal from the pixels themselves which seems to work well enough.

I used it in Storm and I think the original Worms probably used it for the grenade bouncing.

Yep. You could draw your own maps for Worms as bitmap pictures, which means that there are no other options than calculating your normals directly from the bitmap data.

The shots at the top of the thread appear to be done the Mario way as mentioned earlier.

I use multi-tile painting for large tiled objects. Just allow box selection within tile picker so you can select large areas from your tile set.

I usually try to separate the collision map from the tile map so the tiles are only cosmetic. The following images are from my platform engine's editor. The first shows both the collision map and the tile map, the second shows only the collision map.
{"name":"589869","src":"\/\/djungxnpq2nug.cloudfront.net\/image\/cache\/f\/d\/fd9f8d9cb913b5cdd4aea3c08caf7e12.png","w":500,"h":500,"tn":"\/\/djungxnpq2nug.cloudfront.net\/image\/cache\/f\/d\/fd9f8d9cb913b5cdd4aea3c08caf7e12"}
{"name":"589868","src":"\/\/djungxnpq2nug.cloudfront.net\/image\/cache\/4\/d\/4d6594dae3424617649796ea34dca423.png","w":500,"h":500,"tn":"\/\/djungxnpq2nug.cloudfront.net\/image\/cache\/4\/d\/4d6594dae3424617649796ea34dca423"}

This shot is not a good example but the collision map supports slopes of any angle as well as invisible walls/floors or hidden holes. The tedious nature of separate collision/tile maps can be alleviated by adding a prefab tool that can capture both the tiles and their colliders.

The best part of multi-tile tiles is the ability to mix and match within the editor to create new tiles without any extra memory usage. If you use OpenGL then large tiles can even be strippified:).

EDIT: Doh! How do you make the images do the thumbnail thing with the magnify icon?

The forum software does it automatically for all inline images. Whether or not you see a thumbnail is up to your account settings.

The Dizzy games, for example, checked pixel-perfect collision between the character and the scenery. (character had a rectangular bounding box)
See the screenshot in: http://www.yolkfolkserver.co.uk/dta.htm
A very uneven ground, made of a few 8x8 tiles that can be reused many times.

This doesn't prevent the levels from being tile-based : more than 95% of the tiles are empty or full and speed up the check considerably. The pixel-perfect check is only required when the character interacts with a complex tile.

Old techniques (14 years in this case, but I doubt Dizzy was the first) are still useful today for console programmers who deal with limited memory and cpu resources.

I disagree — vector levels provide an increased amount of shape and content variation per kb.

Naturally I feel the need to point towards the Skater Allegro demo (https://svn.sourceforge.net/svnroot/alleg/demos/skater/), which I contributed to. It uses a vector map to produce a level that would be very hard to replicate with tiles, especially in a similar amount of space. At one point it was going to be the next "official" Allegro demo game, but I've no idea what will happen to it now. I'm sure there is an easier way to get it than http access to SVN...

The graphics aren't exactly Castlevania, but a shot (from an older version):

{"name":"DemoGame.png","src":"\/\/djungxnpq2nug.cloudfront.net\/image\/cache\/5\/c\/5c2001210e622cbcc3223606fc1a4894.png","w":720,"h":480,"tn":"\/\/djungxnpq2nug.cloudfront.net\/image\/cache\/5\/c\/5c2001210e622cbcc3223606fc1a4894"}

The key to making tiles look nice is to draw things on blocks of several tiles.
Like in screenshot 2, the lamps were made on 32x32 blocks, though the "original" wall background is only one 16x16 tile.
The bigger the blocks, the less tiled it will look.

The big variety comes from the MANY tilesets used for different parts of the castle. There are completely different graphics for the cellar, gallery, battlements, library, dungeons, etc.