Segment tree can help solve the problem. I've got AC 0.046s 512Kb.
My steps:
1. Read all input and build segment tree
2. Get rmq(i, i + m - 1) for all needed i

I cann't believe. It DON'T get TLE.

    for( long i = m-1; i < (long)arr.size(); ++i )
        cout << *max_element( arr.begin()+i+1-m, arr.begin()+i+1 ) << endl;

=)