PointCNN代码阅读笔记
2018-10-26 本文已影响0人
脑阔不疼
Readme阅读
代码文件内容
- X-conv及PointCNN架构：pointcnn.py
- 分类任务超参数：pointcnn_cls.py
- 分割任务相关：pointcnn_seg.py
关于 X-Conv 和 X-DeConv 的参数
- 以shapenet_x8_2048_fps.py 文件的代码为例
xconv_param_name = ('K', 'D', 'P', 'C', 'links')
xconv_params = [dict(zip(xconv_param_name, xconv_param)) for xconv_param in
[(8, 1, -1, 32 * x, []),
(12, 2, 768, 32 * x, []),
(16, 2, 384, 64 * x, []),
(16, 6, 128, 128 * x, [])]]
xdconv_param_name = ('K', 'D', 'pts_layer_idx', 'qrs_layer_idx')
xdconv_params = [dict(zip(xdconv_param_name, xdconv_param)) for xdconv_param in
[(16, 6, 3, 2),
(12, 6, 2, 1),
(8, 6, 1, 0),
(8, 4, 0, 0)]]
xconv_params
- K：近邻数
- D：膨胀率
- P：输出代表点的数量（-1表示输出输入中的所有点）
- C：输出通道数（深度）
- links：DesNet风格的连接（元组形式，表示该层输入包含倒数X层的输出）
xdconv_params
- K，D：与上同
- pts_layer_idx ：指定某层X-conv的输出作为X-dconv的输入
- qrs_layer_idx ：指定某层X-conv的输出与X-dconv输出混合
PointCNN.py
X-conv结构
def xconv(pts, fts, qrs, tag, N, K, D, P, C, C_pts_fts, is_training, with_X_transformation, depth_multiplier,sorting_method=None, with_global=False):
函数参数
- pts：points
- fts：features
- qrs：queries
- K：近邻数
- D：膨胀率
- P：点数目
- C：通道数（深度）
- with_X_transformation：是否需要X变换
_, indices_dilated = pf.knn_indices_general(qrs, pts, K * D, True)
indices = indices_dilated[:, :, ::D, :]
def knn_indices_general(queries, points, k, sort=True, unique=True):
queries_shape = tf.shape(queries)
batch_size = queries_shape[0]
point_num = queries_shape[1]
D = batch_distance_matrix_general(queries, points)
if unique:
prepare_for_unique_top_k(D, points)
distances, point_indices = tf.nn.top_k(-D, k=k, sorted=sort) # (N, P, K)
batch_indices = tf.tile(tf.reshape(tf.range(batch_size), (-1, 1, 1, 1)), (1, point_num, k, 1))
indices = tf.concat([batch_indices, tf.expand_dims(point_indices, axis=3)], axis=3)
return -distances, indices
def batch_distance_matrix_general(A, B):
r_A = tf.reduce_sum(A * A, axis=2, keep_dims=True)
r_B = tf.reduce_sum(B * B, axis=2, keep_dims=True)
m = tf.matmul(A, tf.transpose(B, perm=(0, 2, 1)))
D = r_A - 2 * m + tf.transpose(r_B, perm=(0, 2, 1))
return D
def prepare_for_unique_top_k(D, A):
indices_duplicated = tf.py_func(find_duplicate_columns, [A], tf.int32)
D += tf.reduce_max(D)*tf.cast(indices_duplicated, tf.float32)
- 涉及维度变化，不太理解
nn_pts = tf.gather_nd(pts, indices, name=tag + 'nn_pts') # (N, P, K, 3)
nn_pts_center = tf.expand_dims(qrs, axis=2, name=tag + 'nn_pts_center') # (N, P, 1, 3)
nn_pts_local = tf.subtract(nn_pts, nn_pts_center, name=tag + 'nn_pts_local') # (N, P, K, 3)
- 得到相对坐标矩阵（第一步）
- nn_pts_local应对应于公式中P'
nn_fts_from_pts_0 = pf.dense(nn_pts_local, C_pts_fts, tag + 'nn_fts_from_pts_0', is_training)
nn_fts_from_pts = pf.dense(nn_fts_from_pts_0, C_pts_fts, tag + 'nn_fts_from_pts', is_training)
if fts is None:
nn_fts_input = nn_fts_from_pts
else:
nn_fts_from_prev = tf.gather_nd(fts, indices, name=tag + 'nn_fts_from_prev')
nn_fts_input = tf.concat([nn_fts_from_pts, nn_fts_from_prev], axis=-1, name=tag + 'nn_fts_input')
处理特征
- 从点云坐标中提取特征（第二步）
- 将提取的特征和原有特征合并（第三步）
if with_X_transformation:
######################## X-transformation #########################
X_0 = pf.conv2d(nn_pts_local, K * K, tag + 'X_0', is_training, (1, K))
X_0_KK = tf.reshape(X_0, (N, P, K, K), name=tag + 'X_0_KK')
X_1 = pf.depthwise_conv2d(X_0_KK, K, tag + 'X_1', is_training, (1, K))
X_1_KK = tf.reshape(X_1, (N, P, K, K), name=tag + 'X_1_KK')
X_2 = pf.depthwise_conv2d(X_1_KK, K, tag + 'X_2', is_training, (1, K), activation=None)
X_2_KK = tf.reshape(X_2, (N, P, K, K), name=tag + 'X_2_KK')
fts_X = tf.matmul(X_2_KK, nn_fts_input, name=tag + 'fts_X')
###################################################################
else:
fts_X = nn_fts_input
X变换矩阵
- 一层卷积+两层深度可分卷积（不理解）
- 卷积核为（1，k），将（N,P,K,3）提升至（N,P,K,K）（第四步）
- 将得到的X变换矩阵与特征矩阵相乘，得到可以直接与卷积核卷积的特征（第五步）
PointCNN架构
train_val_cls.py
- 数据预处理（切片，乱序）
- 训练轮数，batch_size
- 数据增强
- tensorboard
-
准确率与loss
- acc：总准确率
- macc：每类平均准确率
