class ModelCache : public ResourceCache<Model> { //templatized version
(...) //if asks for a texture, call subcache->acquire(x);
TextureCache* subcache;
};

ModelCache cache1;
TextureCache cache2;

cache1.subcache = &cache2;

There is another way of efficiently doing?

Why not make it a parameter of `ModelCache`'s constructor?

I'm not talking about eliminating the dependency but instead if there is another way to "automate" that. Anyway, I'm taking that in consideration because I'm using RAII, and most of the inicialization occurs on the constructor.

You shouldn't be loading textures from within your ResourceCache, it's job is to be a cache. Per the single responsibility principle, you should also have a ResourceLoader that has a reference to your ModelCache and TextureCache. The loader loads the model, and if it needs a texture, it looks in the TextureCache before it tries loading the texture file. When the model is loaded, the loader inserts the model in the ModelCache.

You shouldn't be loading textures from within your ResourceCache, it's job is to be a cache. Per the single responsibility principle, you should also have a ResourceLoader that has a reference to your ModelCache and TextureCache. The loader loads the model, and if it needs a texture, it looks in the TextureCache before it tries loading the texture file. When the model is loaded, the loader inserts the model in the ModelCache.

template <class T> class ResourceCreator {
public:
	virtual shared_ptr<T> create(string handle) = 0;
};

template <class T> class ResourceTable {
public:
	typedef map<string, shared_ptr<T>> ResourceMap;
	
	ResourceTable()
	{
	}

	~ResourceTable()
	{
	}

	shared_ptr<T> acquire(string handle)
	{
		ResourceMap::iterator it = rscmap.find(handle);
		if (it != rscmap.end()) {
			return it->second;
		}
		ResourceCreator<T> creator; //cannot instantiate abstract class
		return creator.create(handle);
	}

	ResourceMap rscmap;
};

The problem is in the commented line. And for that reason I'm thinking to keep the approach that you've posted.

What I've got until now:

template <class T> class ResourceTable {
public:
	typedef map<string, shared_ptr<T>> ResourceMap;

	ResourceTable()
	{
	}

	~ResourceTable()
	{
	}

	T* find(string handle)
	{
		ResourceMap::iterator it = rscmap.find(handle);
		if (it != rscmap.end()) {
			return it->second.get();
		}
		return 0;
	}

	void insert(string handle, T* t) 
        {
		rscmap.insert(make_pair(handle, shared_ptr<T>(t)));
	}

	ResourceMap rscmap;
};

struct A {
};

struct B {
};

class ACreator {
public:
	A* create(string handle, ResourceTable<A>& table1, ResourceTable<B>& table2)
	{
		A* a = table1.find(handle);
		if (a) {
			return a;
		}

		a = new A(); //load

		table1.insert(handle, a);
		return a;
	}
};

int _tmain(int argc, _TCHAR* argv[])
{
	ResourceTable<A> table1;
	ResourceTable<B> table2;

	ACreator ac;
	shared_ptr<A> aptr1(ac.create("foo", table1, table2));
	shared_ptr<A> aptr2(ac.create("foo", table1, table2));
	shared_ptr<A> aptr3(ac.create("foo", table1, table2));
}

Okay. Avoid Resource Managers!

Loading an object becomes such specific thing that the loading process should be handled by themselves and get stored in some

std::map<std::string, YourType> map

YourType* yt = new Type();

yt->Load(stream); //pass a renderer, or a audio map, or something else

map.insert(key, type);

not obeying the SRP? The only class that knows what an object is the class itself, if is not a concrete class should have a loader.

eg.

#1. class Model: you can have a class OBJLoader

#2. class OBJModel: pick it up a Renderer to store the textures, load the data, mtl files and if loads correctly insert that model in a container of your choice.

A Resource Manager can be done, but then you have to subclass a lot of things just for getting a objective: loading specific stuff.