This problem is just to use BFS to keep track of the largest element in each row. The first naive approach comes into my mind it to build a max_list to correspond to the level:
max_list = [-float('inf')] * 10000
max_list[level] = max(max_list[level], node.val)
This is quite intuitive, the problem for this approach is that we don't know the height of the tree, which is not very optimal.
Let's try to improve this, we just build one single temp value cur_max and compare:
cur_max = max(cur_max, node.val)
if level go up, we just result.append(cur_max) then reset cur_max = -float('inf')