For example, lets take a picture of a hill in a basic 2D game.
This would be a hard hill to walk over with normal Byond graphics wouldn't it!
Now lets take a look at how you would be forced to display the mob in this instance.
The black being the logical place to put it, the gray being secondary alternatives. With a pixel based game, you would need to code in the values of each piece of land, or icon, and have the mob dynamically place itself above that, you would be able to form the mob to the ground, with a lot of runtime code, and a lot of values for each piece of land.
Now lets take it if you were able to scan the icons, and figure out what color is where, for instance if there is green piece of land, and there is grass all over it on a variable slope, you would be able to in a list, keep track of which pixels have color, and which don't without the hassle of doing it yourself. For instance, this piece of hill, would be very hard to keep track of, because of its slope.
Now lets go back to the instance before with the hill, and scanning of icons.
You would be able to easily mold the mobs around the hill with little continues coding.
Another good use of this is, the fact that you can have the mob visually sink into the ground, if the ground happens to be sand. You could scan the icon, if the top pixels are within a certain RGB range(Or variable check on the peice of land), it would be considered sand, and the character can visually sink into the sand, and when the sand is over with, he can come out of the sand. Also, you could include raining in a game such as this, where the ground gets wet and muddy, this makes random spots very loose, allowing your character to fall through the ground slightly, while the dynamic ground isn't wet all the way through, the front part of the body could go down slightly more then the back part of the body.