def kdtree_radius_search(root: Node, db: np.ndarray, result_set: RadiusNNResultSet, query: np.ndarray): if root is None: return False if root.is_leaf(): # compare the contents of a leaf leaf_points = db[root.point_indices, :] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) return False # 作业3 # 提示:通过递归的方式实现搜索 # 屏蔽开始 if query[root.axis] <= root.value: kdtree_radius_search(root.left, db, result_set, query) if math.fabs(query[root.axis] - root.value) < result_set.worstDist(): kdtree_radius_search(root.right, db, result_set, query) else: kdtree_radius_search(root.right, db, result_set, query) if math.fabs(query[root.axis] - root.value) < result_set.worstDist(): kdtree_radius_search(root.left, db, result_set, query) # 屏蔽结束 return False
def kdtree_radius_search(root: Node, db: np.ndarray, result_set: RadiusNNResultSet, query: np.ndarray): if root is None: return False if root.is_leaf(): # compare the contents of a leaf leaf_points = db[root.point_indices, :] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) return False # 作业3 # 提示:通过递归的方式实现搜索 # 屏蔽开始 if root.value >= query[root.axis]: if kdtree_knn_search(root.left, db, result_set, query): return True elif math.fabs(root.value - query[root.axis]) < result_set.worst_dist: return kdtree_knn_search(root.right, db, result_set, query) return False # 3、左右子树都不满足,就需要返回上一层树,所以为False else: if kdtree_knn_search(root.right, db, result_set, query): return True elif math.fabs(root.value - query[root.axis]) < result_set.worst_dist: return kdtree_knn_search(root.left, db, result_set, query) return False # 3、左右子树都不满足,就需要返回上一层树,所以为False # 屏蔽结束 return False
def octree_radius_search_fast(root: Octant, db: np.ndarray, result_set: RadiusNNResultSet, query: np.ndarray): if root is None: return False if contains(query, result_set.worstDist(), root): # compare the contents of the octant leaf_points = db[root.point_indices, :] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) # don't need to check any child return False if root.is_leaf and len(root.point_indices) > 0: # compare the contents of a leaf leaf_points = db[root.point_indices, :] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) # check whether we can stop search now return inside(query, result_set.worstDist(), root) # no need to go to most relevant child first, because anyway we will go through all children for c, child in enumerate(root.children): if child is None: continue if False == overlaps(query, result_set.worstDist(), child): continue if octree_radius_search_fast(child, db, result_set, query): return True return inside(query, result_set.worstDist(), root)
def octree_radius_search_fast(root: Octant, db: np.ndarray, result_set: RadiusNNResultSet, query: np.ndarray): if root is None: return False # 1. if query ball contains the octant, no need to check child, just compare all point in it! # 只是不需要向下递归,这里的child是只当前节点的下一层,回溯还是有可能的!所以这里return的false是有原因的嗷! if contains(query, result_set.worstDist(), root): # compare all points: leaf_points = db[root.point_indices, :] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) # no need to check child return False # consider leaf point: if root.is_leaf and len(root.point_indices) > 0: # compare all points: leaf_points = db[root.point_indices, :] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) return inside(query, result_set.worstDist(), root) # 2. check all children for c, child in enumerate(root.children): if child is None: continue if False == overlaps(query, result_set.worstDist(), child): continue if octree_radius_search_fast(child, db, result_set,query): return True return inside(query, result_set.worstDist(), root)
def octree_radius_search(root: Octant, db: np.ndarray, result_set: RadiusNNResultSet, query: np.ndarray): if root is None: return False if root.is_leaf and len(root.point_indices) > 0: # compare the contents of a leaf leaf_points = db[root.point_indices, :] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) # check whether we can stop search now return inside(query, result_set.worstDist(), root) # 作业6 # 屏蔽开始 if root.is_leaf and len(root.point_indices) > 0: # compare the contents of a leaf leaf_points = db[root.point_indices, :] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) # check whether we can stop search now return inside(query, result_set.worstDist(), root) # go to the relevant child first morton_code = 0 if query[0] > root.center[0]: morton_code = morton_code | 1 if query[1] > root.center[1]: morton_code = morton_code | 2 if query[2] > root.center[2]: morton_code = morton_code | 4 if octree_radius_search(root.children[morton_code], db, result_set, query): return True # check other children for c, child in enumerate(root.children): if c == morton_code or child is None: continue if False == overlaps(query, result_set.worstDist(), child): continue if octree_radius_search(child, db, result_set, query): return True # 屏蔽结束 # final check of if we can stop search return inside(query, result_set.worstDist(), root)
def octree_radius_search_plus(root: Octant, db: np.ndarray, result_set: RadiusNNResultSet, query: np.ndarray): if root is None: return False if root.is_leaf and len(root.point_indices) > 0: # compare the contents of a leaf leaf_points = db[root.point_indices, :] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) # check whether we can stop search now return inside(query, result_set.worstDist(), root) # 提前结束:核心-八叉树对3个维度有限制; # 结束条件:其实是两个方向:向下递归,向上回溯 # 1-当一个节点返回True时,表示找到knn了,维度限制导致不会有更优的了,不再向下,立即结束; # 2-最坏距离球如果 inside 当前节点,不需要再向上回溯检查其他节点,立即结束; # 跳过条件:在检查 # 1. search the first relevant child: # 找到最近的孩子,根据查询点的莫顿码 morton_code = 0 if query[0] > root.center[0]: morton_code = morton_code | 1 if query[1] > root.center[1]: morton_code = morton_code | 2 if query[2] > root.center[2]: morton_code = morton_code | 4 if octree_knn_search(root.children[morton_code], db, result_set, query): return True # 2. check other children inv_worstDist = 1 / result_set.worstDist() for c, child in enumerate(root.children): if c == morton_code or child is None: continue if False == overlaps_plus(query, inv_worstDist, child): continue if octree_knn_search(child, db, result_set, query): return True # final check of if we can stop search return inside(query, result_set.worstDist(), root)
def octree_radius_search_fast(root: Octant, db: np.ndarray, result_set: RadiusNNResultSet, query: np.ndarray): if root is None: return False # 作业5 # 提示:尽量利用上面的inside、overlaps、contains等函数 # 屏蔽开始 #如果球包含了这个root Octant则不需要去它的子节点找了 if (root.is_leaf and len(root.point_indices) > 0) or contains( query, result_set.worstDist(), root): # compare the contents of a leaf leaf_points = db[root.point_indices, :] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) # check whether we can stop search now return inside(query, result_set.worstDist(), root) #如果球包含了这个root Octant则不需要去它的子节点找了 children_idx = 0 if query[0] > root.center[0]: # x轴 children_idx = children_idx | 1 if query[1] > root.center[1]: # y轴 children_idx = children_idx | 2 if query[2] > root.center[2]: # z轴 children_idx = children_idx | 4 # 如果在这个子octant中发现,查询点在该子octant中,同时由worst_dist构成的球也被octant包含,所以直接返回 if octree_knn_search(root.children[children_idx], db, result_set, query): return True # 如果不满足上边的情况则需要遍历其他子octant for c, child in enumerate(root.children): if c == children_idx or root.children[c]: continue if overlaps(query, result_set.worstDist(), child) == False: continue if octree_knn_search(root.children[c], db, result_set, query): return True # 屏蔽结束 return inside(query, result_set.worstDist(), root)
def octree_radius_search(root: Octant, db: np.ndarray, result_set: RadiusNNResultSet, query: np.ndarray): if root is None: return False if root.is_leaf and len(root.point_indices) > 0: # compare the contents of a leaf leaf_points = db[root.point_indices, :] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) # check whether we can stop search now return inside(query, result_set.worstDist(), root) # 作业6 # 屏蔽开始 # 屏蔽结束 # final check of if we can stop search return inside(query, result_set.worstDist(), root)
def kdtree_radius_search(root: Node, db: np.ndarray, result_set: RadiusNNResultSet, query: np.ndarray): if root is None: return False if root.is_leaf(): # compare the contents of a leaf leaf_points = db[root.point_indices, :] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) return False # 作业3 # 提示:通过递归的方式实现搜索 # 屏蔽开始 # 屏蔽结束 return False
def radius_search(root: Node, result_set: RadiusNNResultSet, key): if root is None: return False # compare the root itself result_set.add_point(math.fabs(root.key - key), root.value) if root.key >= key: # iterate left branch first if radius_search(root.left, result_set, key): return True elif math.fabs(root.key-key) < result_set.worstDist(): return radius_search(root.right, result_set, key) return False else: # iterate right branch first if radius_search(root.right, result_set, key): return True elif math.fabs(root.key-key) < result_set.worstDist(): return radius_search(root.left, result_set, key) return False
def octree_radius_search_fast(root: Octant, db: np.ndarray, result_set: RadiusNNResultSet, query: np.ndarray): if root is None: return False # 作业5 # 提示:尽量利用上面的inside、overlaps、contains等函数 # 屏蔽开始 if contains(query, result_set.worstDist(), root): # the octant is contained by the query ball with resule_set.worstDist() leaf_points = db[root.point_indices, :] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) # check whether we can stop search now return False if root.is_leaf and len(root.point_indices) > 0: # compare the contents of a leaf leaf_points = db[root.point_indices, :] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) # check whether we can stop search now return inside(query, result_set.worstDist(), root) # check other children for c, child in enumerate(root.children): if child is None: continue # if an octant is not overlapping with query ball, skip if overlaps(query, result_set.worstDist(), child) == False: continue if octree_radius_search(child, db, result_set, query): return True # 屏蔽结束 return inside(query, result_set.worstDist(), root)
def kdtree_radius_search(root: Node, db: np.ndarray, result_set: RadiusNNResultSet, query: np.ndarray): if root is None: return if root.is_leaf(): leaf_points = db[root.point_indices] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) return if query[root.axis] <= root.value: kdtree_radius_search(root.left, db, result_set, query) if math.fabs(query[root.axis] - root.value) < result_set.get_worst_dist(): kdtree_radius_search(root.right, db, result_set, query) else: kdtree_radius_search(root.right, db, result_set, query) if math.fabs(query[root.axis] - root.value) < result_set.get_worst_dist(): kdtree_radius_search(root.left, db, result_set, query)
def main(): # configuration leaf_size = 32 min_extent = 0.0001 k = 8 radius = 1 #root_dir = '/Users/renqian/cloud_lesson/kitti' # 数据集路径 #cat = os.listdir(root_dir) #iteration_num = len(cat) # load date filename = '/home/ljn/SLAM/dateset/000000.bin' db_np = read_velodyne_bin(filename) iteration = 1 print("octree -----------------------------------") construction_time_sum = 0 knn_time_sum = 0 radius_time_sum = 0 brute_time_sum = 0 result_set_knn = KNNResultSet(capacity=k) query = db_np[95, :] begin_t = time.time() root = octree.octree_construction(db_np, leaf_size, min_extent) construction_time_sum += time.time() - begin_t begin_t = time.time() octree.octree_knn_search(root, db_np, result_set_knn, query) knn_time_sum += time.time() - begin_t print('knn search result\n', result_set_knn) begin_t = time.time() result_set_rnn = RadiusNNResultSet(radius=radius) octree.octree_radius_search_fast(root, db_np, result_set_rnn, query) radius_time_sum += time.time() - begin_t print('rnn search result\n', result_set_rnn) begin_t = time.time() diff = np.linalg.norm(np.expand_dims(query, 0) - db_np, axis=1) nn_idx = np.argsort(diff) nn_dist = diff[nn_idx] nn_dist_idx_pre = np.linspace(0, nn_dist.shape[0], nn_dist.shape[0]) nn_dist_idx = nn_dist_idx_pre[nn_idx] brute_time_sum += time.time() - begin_t #brute knn search result_set_knn_brute = KNNResultSet(capacity=k) for index in range(k): result_set_knn_brute.add_point(nn_dist[index], nn_dist_idx[index]) # brute radiusNN search result_set_rnn_brute = RadiusNNResultSet(radius=radius) for index in range(nn_dist.shape[0]): if nn_dist[index] < radius: result_set_rnn_brute.add_point(nn_dist[index], nn_dist_idx[index]) continue else: break print("Octree: build %.3f, knn %.3f, radius %.3f, brute %.3f" % (construction_time_sum * 1000, knn_time_sum * 1000, radius_time_sum * 1000, brute_time_sum * 1000)) print("kdtree -----------------------------------") construction_time_sum = 0 knn_time_sum = 0 radius_time_sum = 0 brute_time_sum = 0 for i in range(iteration): query = db_np[95, :] begin_t = time.time() root = kdtree.kdtree_construction(db_np, leaf_size) construction_time_sum += time.time() - begin_t begin_t = time.time() result_set_knn = KNNResultSet(capacity=k) kdtree.kdtree_knn_search(root, db_np, result_set_knn, query) knn_time_sum += time.time() - begin_t begin_t = time.time() result_set_rnn = RadiusNNResultSet(radius=radius) kdtree.kdtree_radius_search(root, db_np, result_set_rnn, query) radius_time_sum += time.time() - begin_t begin_t = time.time() diff = np.linalg.norm(np.expand_dims(query, 0) - db_np, axis=1) nn_idx = np.argsort(diff) nn_dist = diff[nn_idx] brute_time_sum += time.time() - begin_t nn_dist_idx_pre = np.linspace(0, nn_dist.shape[0] - 1, nn_dist.shape[0]) nn_dist_idx = nn_dist_idx_pre[nn_idx] # brute knn search result_set_knn_brute = KNNResultSet(capacity=k) for index in range(k): result_set_knn_brute.add_point(nn_dist[index], nn_dist_idx[index]) # brute radiusNN search result_set_rnn_brute = RadiusNNResultSet(radius=radius) for index in range(nn_dist.shape[0]): if nn_dist[index] < radius: result_set_rnn_brute.add_point(nn_dist[index], nn_dist_idx[index]) continue else: break print("Kdtree: build %.3f, knn %.3f, radius %.3f, brute %.3f" % (construction_time_sum * 1000, knn_time_sum * 1000, radius_time_sum * 1000, brute_time_sum * 1000)) print("scipy kdtree -----------------------------------") construction_time_sum = 0 knn_time_sum = 0 radius_time_sum = 0 query = db_np[95, :] begin_t = time.time() tree = spatial.KDTree(db_np, leaf_size) construction_time_sum += time.time() - begin_t #no knn begin_t = time.time() knn_time_sum += time.time() - begin_t begin_t = time.time() result_set_rnn = tree.query_ball_point(query, radius) radius_time_sum += time.time() - begin_t print('rnn search result\n', result_set_rnn) print("Octree: build %.3f, knn %.3f, radius %.3f" % (construction_time_sum * 1000, knn_time_sum * 1000, radius_time_sum * 1000))
def octree_radius_search(root: Octant, db: np.ndarray, result_set: RadiusNNResultSet, query: np.ndarray, search='normal'): if root is None: return False # 作业5 # 另外一种极端情况:查询半径已经包括了整个区域 if search == 'fast': if contains(query, result_set.worstDist(), root): leaf_points = db[root.point_indices, :] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) return False # 如果是叶子,则将该区域内返回到result_set if root.is_leaf and len(root.point_indices) > 0: # compare the contents of a leaf leaf_points = db[root.point_indices, :] diff = np.linalg.norm(np.expand_dims(query, 0) - leaf_points, axis=1) for i in range(diff.shape[0]): result_set.add_point(diff[i], root.point_indices[i]) # check whether we can stop search now return inside(query, result_set.worstDist(), root) # 作业6 # 屏蔽开始 child_idx = -1 if search == 'normal': child_idx = 0 # 判断所查询点在八叉树的位置 if query[0] > root.center[0]: child_idx = child_idx | 1 if query[1] > root.center[1]: child_idx = child_idx | 2 if query[2] > root.center[2]: child_idx = child_idx | 4 # 递归判断是否在该区域就可以找到足够多的点 if octree_radius_search(root.children[child_idx], db, result_set, query, search): return True # 没有在查询点区域找到,对其他区域进行搜索 for c, child in enumerate(root.children): if c == child_idx or child is None: # 搜索区域没有点或所搜索到查询点区域,skip continue if False == overlaps(query, result_set.worstDist(), child): # 搜索区域与查询点和最坏距离构成的球面没有交点,skip continue if octree_radius_search(child, db, result_set, query, search): # 其他情况可以进入搜索区域搜索(递归) return True # 屏蔽结束 # final check of if we can stop search return inside(query, result_set.worstDist(), root)