我不太了解wikipedia中的O(log n)最近邻算法。
- ...
- ...
- 算法展开树的递归,在每个节点执行以下步骤:
醇>
- ...
- 该算法检查在分割平面的另一侧是否可能存在比当前最佳点更接近搜索点的任何点。在概念上,这通过使分裂超平面与搜索点周围的超球面相交来完成,该超球面具有等于当前最近距离的半径。由于超平面都是轴对齐的,因此将其实现为简单的比较,以查看搜索点的分割坐标与当前节点之间的差异是否小于从搜索点到当前最佳的距离(总坐标)。
- 如果超球面穿过平面,则平面另一侧可能有更近的点,因此算法必须从当前节点向下移动树的另一个分支,寻找更近的点,遵循相同的递归过程。整个搜索。
- 如果超球面不与分裂平面相交,则算法继续向上走树,并消除该节点另一侧的整个分支。
3.2令我感到困惑,我看到了this个问题。我正在用Java实现算法,不确定我是否正确。
//Search children branches, if axis aligned distance is less than current distance
if (node.lesser!=null) {
KdNode lesser = node.lesser;
int axis = lesser.depth % lesser.k;
double axisAlignedDistance = Double.MAX_VALUE;
if (axis==X_AXIS) axisAlignedDistance = Math.abs(lastNode.id.x-lesser.id.x);
if (axis==Y_AXIS) axisAlignedDistance = Math.abs(lastNode.id.y-lesser.id.y);
else if (axis==Z_AXIS) axisAlignedDistance = Math.abs(lastNode.id.z-lesser.id.z);
//Continue down lesser branch
if (axisAlignedDistance<=lastDistance && !set.contains(lesser)) {
searchNode(value,lesser,set,K);
}
}
if (node.greater!=null) {
KdNode greater = node.greater;
int axis = greater.depth % greater.k;
double axisAlignedDistance = Double.MAX_VALUE;
if (axis==X_AXIS) axisAlignedDistance = Math.abs(lastNode.id.x-greater.id.x);
if (axis==Y_AXIS) axisAlignedDistance = Math.abs(lastNode.id.y-greater.id.y);
else if (axis==Z_AXIS)axisAlignedDistance = Math.abs(lastNode.id.z-greater.id.z);
//Continue down greater branch
if (axisAlignedDistance<=lastDistance && !set.contains(greater)) {
searchNode(value,greater,set,K);
}
}
上面的代码是否完成了算法的3.2方面?特别是我填充“axisAlignedDistance”变量。
您可以找到KDTree here的完整源代码。
感谢您提供任何帮助/指示。
答案 0 :(得分:1)
我正在添加这个,希望它可以帮助其他人找到同样的问题。我最后用以下代码解决了3.2问题。虽然,我不确定它是否100%正确。它已通过我提出的所有测试。上面的原始代码在许多相同的测试用例中都失败了。
使用Point,Line,Rectangle和Cube对象的更明确的解决方案:
int axis = node.depth % node.k;
KdNode lesser = node.lesser;
KdNode greater = node.greater;
//Search children branches, if axis aligned distance is less than current distance
if (lesser!=null && !examined.contains(lesser)) {
examined.add(lesser);
boolean lineIntersectsRect = false;
Line line = null;
Cube cube = null;
if (axis==X_AXIS) {
line = new Line(new Point(value.x-lastDistance,value.y,value.z), new Point(value.x+lastDistance,value.y,value.z));
Point tul = new Point(Double.NEGATIVE_INFINITY,Double.NEGATIVE_INFINITY,Double.NEGATIVE_INFINITY);
Point tur = new Point(node.id.x,Double.NEGATIVE_INFINITY,Double.NEGATIVE_INFINITY);
Point tlr = new Point(node.id.x,Double.POSITIVE_INFINITY,Double.NEGATIVE_INFINITY);
Point tll = new Point(Double.NEGATIVE_INFINITY,Double.POSITIVE_INFINITY,Double.NEGATIVE_INFINITY);
Rectangle trect = new Rectangle(tul,tur,tlr,tll);
Point bul = new Point(Double.NEGATIVE_INFINITY,Double.NEGATIVE_INFINITY,Double.POSITIVE_INFINITY);
Point bur = new Point(node.id.x,Double.NEGATIVE_INFINITY,Double.POSITIVE_INFINITY);
Point blr = new Point(node.id.x,Double.POSITIVE_INFINITY,Double.POSITIVE_INFINITY);
Point bll = new Point(Double.NEGATIVE_INFINITY,Double.POSITIVE_INFINITY,Double.POSITIVE_INFINITY);
Rectangle brect = new Rectangle(bul,bur,blr,bll);
cube = new Cube(trect,brect);
lineIntersectsRect = cube.inserects(line);
} else if (axis==Y_AXIS) {
line = new Line(new Point(value.x,value.y-lastDistance,value.z), new Point(value.x,value.y+lastDistance,value.z));
Point tul = new Point(Double.NEGATIVE_INFINITY,Double.NEGATIVE_INFINITY,Double.NEGATIVE_INFINITY);
Point tur = new Point(Double.POSITIVE_INFINITY,Double.NEGATIVE_INFINITY,Double.NEGATIVE_INFINITY);
Point tlr = new Point(Double.POSITIVE_INFINITY,node.id.y,Double.NEGATIVE_INFINITY);
Point tll = new Point(Double.NEGATIVE_INFINITY,node.id.y,Double.NEGATIVE_INFINITY);
Rectangle trect = new Rectangle(tul,tur,tlr,tll);
Point bul = new Point(Double.NEGATIVE_INFINITY,Double.NEGATIVE_INFINITY,Double.POSITIVE_INFINITY);
Point bur = new Point(Double.POSITIVE_INFINITY,Double.NEGATIVE_INFINITY,Double.POSITIVE_INFINITY);
Point blr = new Point(Double.POSITIVE_INFINITY,node.id.y,Double.POSITIVE_INFINITY);
Point bll = new Point(Double.NEGATIVE_INFINITY,node.id.y,Double.POSITIVE_INFINITY);
Rectangle brect = new Rectangle(bul,bur,blr,bll);
cube = new Cube(trect,brect);
lineIntersectsRect = cube.inserects(line);
} else {
line = new Line(new Point(value.x,value.y,value.z-lastDistance), new Point(value.x,value.y,value.z+lastDistance));
Point tul = new Point(Double.NEGATIVE_INFINITY,Double.NEGATIVE_INFINITY,Double.NEGATIVE_INFINITY);
Point tur = new Point(Double.POSITIVE_INFINITY,Double.NEGATIVE_INFINITY,Double.NEGATIVE_INFINITY);
Point tlr = new Point(Double.POSITIVE_INFINITY,Double.POSITIVE_INFINITY,Double.NEGATIVE_INFINITY);
Point tll = new Point(Double.NEGATIVE_INFINITY,Double.POSITIVE_INFINITY,Double.NEGATIVE_INFINITY);
Rectangle trect = new Rectangle(tul,tur,tlr,tll);
Point bul = new Point(Double.NEGATIVE_INFINITY,Double.NEGATIVE_INFINITY,node.id.z);
Point bur = new Point(Double.POSITIVE_INFINITY,Double.NEGATIVE_INFINITY,node.id.z);
Point blr = new Point(Double.POSITIVE_INFINITY,Double.POSITIVE_INFINITY,node.id.z);
Point bll = new Point(Double.NEGATIVE_INFINITY,Double.POSITIVE_INFINITY,node.id.z);
Rectangle brect = new Rectangle(bul,bur,blr,bll);
cube = new Cube(trect,brect);
lineIntersectsRect = cube.inserects(line);
}
//Continue down lesser branch
if (lineIntersectsRect) {
searchNode(value,lesser,K,results,examined);
}
}
if (greater!=null && !examined.contains(greater)) {
examined.add(greater);
boolean lineIntersectsRect = false;
Line line = null;
Cube cube = null;
if (axis==X_AXIS) {
line = new Line(new Point(value.x-lastDistance,value.y,value.z), new Point(value.x+lastDistance,value.y,value.z));
Point tul = new Point(node.id.x,Double.NEGATIVE_INFINITY,Double.NEGATIVE_INFINITY);
Point tur = new Point(Double.POSITIVE_INFINITY,Double.NEGATIVE_INFINITY,Double.NEGATIVE_INFINITY);
Point tlr = new Point(Double.POSITIVE_INFINITY,Double.POSITIVE_INFINITY,Double.NEGATIVE_INFINITY);
Point tll = new Point(node.id.x,Double.POSITIVE_INFINITY,Double.NEGATIVE_INFINITY);
Rectangle trect = new Rectangle(tul,tur,tlr,tll);
Point bul = new Point(node.id.x,Double.NEGATIVE_INFINITY,Double.POSITIVE_INFINITY);
Point bur = new Point(Double.POSITIVE_INFINITY,Double.NEGATIVE_INFINITY,Double.POSITIVE_INFINITY);
Point blr = new Point(Double.POSITIVE_INFINITY,Double.POSITIVE_INFINITY,Double.POSITIVE_INFINITY);
Point bll = new Point(node.id.x,Double.POSITIVE_INFINITY,Double.POSITIVE_INFINITY);
Rectangle brect = new Rectangle(bul,bur,blr,bll);
cube = new Cube(trect,brect);
lineIntersectsRect = cube.inserects(line);
} else if (axis==Y_AXIS) {
line = new Line(new Point(value.x,value.y-lastDistance,value.z), new Point(value.x,value.y+lastDistance,value.z));
Point tul = new Point(Double.NEGATIVE_INFINITY,node.id.y,Double.NEGATIVE_INFINITY);
Point tur = new Point(Double.POSITIVE_INFINITY,node.id.y,Double.NEGATIVE_INFINITY);
Point tlr = new Point(Double.POSITIVE_INFINITY,Double.POSITIVE_INFINITY,Double.NEGATIVE_INFINITY);
Point tll = new Point(Double.NEGATIVE_INFINITY,Double.POSITIVE_INFINITY,Double.NEGATIVE_INFINITY);
Rectangle trect = new Rectangle(tul,tur,tlr,tll);
Point bul = new Point(Double.NEGATIVE_INFINITY,node.id.y,Double.POSITIVE_INFINITY);
Point bur = new Point(Double.POSITIVE_INFINITY,node.id.y,Double.POSITIVE_INFINITY);
Point blr = new Point(Double.POSITIVE_INFINITY,Double.POSITIVE_INFINITY,Double.POSITIVE_INFINITY);
Point bll = new Point(Double.NEGATIVE_INFINITY,Double.POSITIVE_INFINITY,Double.POSITIVE_INFINITY);
Rectangle brect = new Rectangle(bul,bur,blr,bll);
cube = new Cube(trect,brect);
lineIntersectsRect = cube.inserects(line);
} else {
line = new Line(new Point(value.x,value.y,value.z-lastDistance), new Point(value.x,value.y,value.z+lastDistance));
Point tul = new Point(Double.NEGATIVE_INFINITY,Double.NEGATIVE_INFINITY,node.id.z);
Point tur = new Point(Double.POSITIVE_INFINITY,Double.NEGATIVE_INFINITY,node.id.z);
Point tlr = new Point(Double.POSITIVE_INFINITY,Double.POSITIVE_INFINITY,node.id.z);
Point tll = new Point(Double.NEGATIVE_INFINITY,Double.POSITIVE_INFINITY,node.id.z);
Rectangle trect = new Rectangle(tul,tur,tlr,tll);
Point bul = new Point(Double.NEGATIVE_INFINITY,Double.NEGATIVE_INFINITY,Double.POSITIVE_INFINITY);
Point bur = new Point(Double.POSITIVE_INFINITY,Double.NEGATIVE_INFINITY,Double.POSITIVE_INFINITY);
Point blr = new Point(Double.POSITIVE_INFINITY,Double.POSITIVE_INFINITY,Double.POSITIVE_INFINITY);
Point bll = new Point(Double.NEGATIVE_INFINITY,Double.POSITIVE_INFINITY,Double.POSITIVE_INFINITY);
Rectangle brect = new Rectangle(bul,bur,blr,bll);
cube = new Cube(trect,brect);
lineIntersectsRect = cube.inserects(line);
}
//Continue down greater branch
if (lineIntersectsRect) {
searchNode(value,greater,K,results,examined);
}
}
我认为this更简单的代码也应该有效,它已经通过了与上述代码相同的所有测试。
int axis = node.depth % node.k;
KdNode lesser = node.lesser;
KdNode greater = node.greater;
//Search children branches, if axis aligned distance is less than current distance
if (lesser!=null && !examined.contains(lesser)) {
examined.add(lesser);
double p1 = Double.MIN_VALUE;
double p2 = Double.MIN_VALUE;
if (axis==X_AXIS) {
p1 = node.id.x;
p2 = value.x-lastDistance;
} else if (axis==Y_AXIS) {
p1 = node.id.y;
p2 = value.y-lastDistance;
} else {
p1 = node.id.z;
p2 = value.z-lastDistance;
}
boolean lineIntersectsCube = ((p2<=p1)?true:false);
//Continue down lesser branch
if (lineIntersectsCube) {
searchNode(value,lesser,K,results,examined);
}
}
if (greater!=null && !examined.contains(greater)) {
examined.add(greater);
double p1 = Double.MIN_VALUE;
double p2 = Double.MIN_VALUE;
if (axis==X_AXIS) {
p1 = node.id.x;
p2 = value.x+lastDistance;
} else if (axis==Y_AXIS) {
p1 = node.id.y;
p2 = value.y+lastDistance;
} else {
p1 = node.id.z;
p2 = value.z+lastDistance;
}
boolean lineIntersectsCube = ((p2>=p1)?true:false);
//Continue down greater branch
if (lineIntersectsCube) {
searchNode(value,greater,K,results,examined);
}
}
答案 1 :(得分:0)
重要的是要注意以下3:
&#34; 算法展开树的递归,执行 在每个节点执行以下步骤:&#34;
您的实现似乎只是进行递归调用,而不是在递归递归时进行任何操作。
虽然看起来您正在检测交叉点,但是您没有执行退出(退出)递归的步骤。在进行递归调用之后,需要执行0个步骤(包括3.2)。
3.2状态:&#34;如果超球面不与分裂平面相交,那么 算法继续向上走树,并在整个分支上 该节点的另一面被淘汰&#34;
这是什么意思?到达基本情况(算法到达叶节点)后,递归开始展开。在解开每个递归级别之后,算法检查子树是否可能包含更近的邻居。如果可以,则在该子树上进行另一次递归调用,否则算法将继续展开(在树上行走)。
你应该从这开始:
&#34;从根节点开始,算法向下移动树 递归地,以与搜索点相同的方式递归 插入&#34;
然后开始考虑在递归的过程中采取什么步骤以及何时采取措施。
这是一个具有挑战性的算法。如果您已完成此数据结构的insert()方法,则可以将其用作此算法的开始框架。
修改强>
//Used to not re-examine nodes
Set<KdNode> examined = new HashSet<KdNode>();
它可能更容易,更快,并且可以使用更少的内存来简单地在每个节点中放置一个标记,您可以将其标记为&#34;已访问&#34;。理想情况下,HashSets有一个恒定的时间查找,但实际上没有办法保证这一点。这样,您不必在每次访问时搜索一组节点。在搜索树之前,您可以在O(N)
时间内将所有这些值初始化为false。
答案 2 :(得分:0)
是的,维基百科上的KD树中的NN(最近邻)搜索的描述有点难以理解。在NN KD Tree搜索中获得大量顶级Google搜索结果是完全错误的,这无济于事!
请参阅https://stackoverflow.com/a/37107030/591720以获取正确的解释/算法。