1. 参考文献
3D Fully Convolutional Network for Vehicle Detection in Point Cloud
2. 模型实现
''' Baidu Inc. Ref: 3D Fully Convolutional Network for Vehicle Detection in Point Cloud Author: HSW Date: 2018-05-02 ''' import sys import numpy as np import tensorflow as tf from prepare_data2 import * from baidu_cnn_3d import * KITTI_TRAIN_DATA_CNT = 7481 KITTI_TEST_DATA_CNT = 7518 # create 3D-CNN Model def create_graph(sess, modelType = 0, voxel_shape = (400, 400, 20), activation=tf.nn.relu, is_train = True): ''' Inputs: sess: tensorflow Session Object voxel_shape: voxel shape for network first layer activation: phrase_train: Outputs: voxel, graph, sess ''' voxel = tf.placeholder(tf.float32, [None, voxel_shape[0], voxel_shape[1], voxel_shape[2], 1]) phase_train = tf.placeholder(tf.bool, name="phase_train") if is_train else None with tf.variable_scope("3D_CNN_Model") as scope: model = Full_CNN_3D_Model() model.cnn3d_graph(voxel, modelType = modelType, activation=activation, phase_train = is_train) if is_train: initialized_var = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope="3D_CNN_model") sess.run(tf.variables_initializer(initialized_var)) return voxel, model, phase_train # read batch data def read_batch_data(batch_size, data_set_dir,objectType = "Car", split = "training", resolution=(0.2, 0.2, 0.2), scale=0.25, limitX = (0,80), limitY=(-40,40), limitZ=(-2.5,1.5)): ''' Inputs: batch_size: data_set_dir: objectType: default is "Car" split: default is "training" resolution: scale: outputSize / inputSize limitX: limitY: limitZ: Outputs: ''' kitti_3DVoxel = kitti_3DVoxel_interface(data_set_dir, objectType = objectType, split=split, scale = scale, resolution = resolution, limitX = limitX, limitY = limitY, limitZ = limitZ) TRAIN_PROCESSED_IDX = 0 TEST_PROCESSED_IDX = 0 if split == "training": while TRAIN_PROCESSED_IDX < KITTI_TRAIN_DATA_CNT: batch_voxel = [] batch_g_obj = [] batch_g_cord = [] idx = 0 while idx < batch_size and TRAIN_PROCESSED_IDX < KITTI_TRAIN_DATA_CNT: print(TRAIN_PROCESSED_IDX) voxel, g_obj, g_cord = kitti_3DVoxel.read_kitti_data(TRAIN_PROCESSED_IDX) TRAIN_PROCESSED_IDX += 1 if voxel is None: continue idx += 1 # print(voxel.shape) batch_voxel.append(voxel) batch_g_obj.append(g_obj) batch_g_cord.append(g_cord) yield np.array(batch_voxel, dtype=np.float32)[:, :, :, :, np.newaxis], np.array(batch_g_obj, dtype=np.float32), np.array(batch_g_cord, dtype=np.float32) elif split == "testing": while TEST_PROCESSED_IDX < KITTI_TEST_DATA_CNT: batch_voxel = [] idx = 0 while idx < batch_size and TEST_PROCESSED_IDX < KITTI_TEST_DATA_CNT: voxel = kitti_3DVoxel.read_kitti_data(iter * batch_size + idx) TEST_PROCESSED_IDX += 1 if voxel is None: continue idx += 1 batch_voxel.append(voxel) yield np.array(batch_voxel, dtype=np.float32)[:, :, :, :, np.newaxis] # train 3D-CNN Model def train(batch_num, data_set_dir, modelType = 0, objectType = "Car", resolution=(0.2,0.2,0.2), scale = 0.25, lr=0.01, limitX=(0,80), limitY=(-40,40), limitZ=(-2.5,1.5), epoch=101): ''' Inputs: batch_num: data_set_dir: modelType: objectType: resolution: scale: lr: limitX, limitY, limitZ: Outputs: None ''' batch_size = batch_num training_epochs = epoch sizeX = int(round((limitX[1] - limitX[0]) / resolution[0])) sizeY = int(round((limitY[1] - limitY[0]) / resolution[0])) sizeZ = int(round((limitZ[1] - limitZ[0]) / resolution[0])) voxel_shape = (sizeX, sizeY, sizeZ) with tf.Session() as sess: voxel, model, phase_train = create_graph(sess, modelType = modelType, voxel_shape = voxel_shape, activation=tf.nn.relu, is_train = True) saver = tf.train.Saver() total_loss, obj_loss, cord_loss, is_obj_loss, non_obj_loss, g_obj, g_cord, y_pred = model.loss_Fun(lossType = 0, cord_loss_weight = 0.02) optimizer = model.create_optimizer(total_loss, optType = "Adam", learnRate = 0.001) init = tf.global_variables_initializer() sess.run(init) for epoch in range(training_epochs): batchCnt = 0; for (batch_voxel, batch_g_obj, batch_g_cord) in read_batch_data(batch_size, data_set_dir, objectType = objectType, split = "training", resolution = resolution, scale = scale, limitX = limitX, limitY = limitY, limitZ = limitZ): # print("batch_g_obj") # print(batch_g_obj.shape) sess.run(optimizer, feed_dict={voxel: batch_voxel, g_obj: batch_g_obj, g_cord: batch_g_cord, phase_train: True}) cord_cost = sess.run(cord_loss, feed_dict={voxel: batch_voxel, g_obj: batch_g_obj, g_cord: batch_g_cord, phase_train: True}) obj_cost = sess.run(is_obj_loss, feed_dict={voxel: batch_voxel, g_obj: batch_g_obj, g_cord: batch_g_cord, phase_train: True}) non_obj_cost = sess.run(non_obj_loss, feed_dict={voxel: batch_voxel, g_obj: batch_g_obj, g_cord: batch_g_cord, phase_train: True}) print("Epoch: ", (epoch + 1), ",", "BatchNum: ", (batchCnt + 1), "," , "cord_cost = ", "{:.9f}".format(cord_cost)) print("Epoch: ", (epoch + 1), ",", "BatchNum: ", (batchCnt + 1), "," , "obj_cost = ", "{:.9f}".format(obj_cost)) print("Epoch: ", (epoch + 1), ",", "BatchNum: ", (batchCnt + 1), "," , "non_obj_cost = ", "{:.9f}".format(non_obj_cost)) batchCnt += 1 if (epoch > 0) and (epoch % 10 == 0): saver.save(sess, "velodyne_kitti_train_" + str(epoch) + ".ckpt") print("Training Finishied !") # test 3D-CNN Model def test(batch_num, data_set_dir, modelType = 0, objectType = "Car", resolution=(0.2, 0.2, 0.2), scale = 0.25, limitX = (0, 80), limitY = (-40, 40), limitZ=(-2.5, 1.5)): ''' Inputs: batch_num: data_set_dir: resolution: scale: limitX, limitY, limitZ: Outputs: None ''' sizeX = int(round((limitX[1] - limitX[0]) / resolution[0])) sizeY = int(round((limitY[1] - limitY[0]) / resolution[0])) sizeZ = int(round((limitZ[1] - limitZ[0]) / resolution[0])) voxel_shape = (sizeX, sizeY, sizeZ) batch_size = batch_num; batch_voxel = read_batch_data(batch_num, data_set_dir, objectType = objectType, split="Testing", resolution=resolution, scale=scale, limitX=limitX, limitY=limitY, limitZ=limitZ) batch_voxel_x = batch_voxel.reshape(1, batch_voxel.shape[0], batch_voxel.shape[1], batch_voxel.shape[2], 1) with tf.Session() as sess: is_train = False voxel, model, phase_train = create_graph(sess, modelType = modelType, voxel_shape = voxel_shape, activation=tf.nn.relu, is_train = False) new_saver = tf.train.import_meta_graph("velodyne_kitti_train_40.ckpt.meta") last_model = "./velodyne_kitti_train_40.ckpt" saver.restore(sess, last_model) objectness = model.objectness cordinate = model.cordinate y_pred = model.y objectness = sess.run(objectness, feed_dict={voxel: batch_voxel_x})[0, :, :, :, 0] cordinate = sess.run(cordinate, feed_dict={voxel:batch_voxel_x})[0] y_pred = sess.run(y_pred, feed_dict={voxel: batch_voxel_x})[0, :, :, :, 0] idx = np.where(y_pred >= 0.995) spheres = np.vstack((index[0], np.vstack((index[1], index[2])))).transpose() centers = spheres_to_centers(spheres, scale = scale, resolution=resolution, limitX = limitX, limitY = limitY, limitZ = limitZ) corners = cordinate[idx].reshape[-1, 8, 3] + centers[:, np.newaxis] print(centers) print(corners) if __name__ == "__main__": batch_num = 3 data_set_dir = "/home/hsw/桌面/PCL_API_Doc/frustum-pointnets-master/dataset" modelType = 1 objectType = "Car" resolution = (0.2, 0.2, 0.2) scale = 0.25 lr = 0.001 limitX = (0, 80) limitY = (-40, 40) limitZ = (-2.5, 1.5) epoch = 101 train(batch_num, data_set_dir = data_set_dir, modelType = modelType, objectType = objectType, resolution=resolution, scale=scale, lr =lr, limitX = limitX, limitY = limitY, limitZ = limitZ) saver = tf.train.Saver()2.1 网络模型
''' Baidu Inc. Ref: 3D Fully Convolutional Network for Vehicle Detection in Point Cloud Author: HSW Date: 2018-05-02 ''' import numpy as np import tensorflow as tf class Full_CNN_3D_Model(object): ''' Define Full CNN Model ''' def __init__(self): pass; def cnn3d_graph(self, voxel, modelType = 0, activation = tf.nn.relu, phase_train = True): if modelType == 0: # Modefied 3D-CNN, 该网络结构不可使用,因为降采样太严重(降采样1/8)导致在预测时会出现较大误差 self.layer1 = self.conv3d_layer(voxel , 1, 16, 5, 5, 5, [1, 2, 2, 2, 1], name="layer1", activation=activation, phase_train=phase_train) self.layer2 = self.conv3d_layer(self.layer1, 16, 32, 5, 5, 5, [1, 2, 2, 2, 1], name="layer2", activation=activation, phase_train=phase_train) self.layer3 = self.conv3d_layer(self.layer2, 32, 64, 3, 3, 3, [1, 2, 2, 2, 1], name="layer3", activation=activation, phase_train=phase_train) self.layer4 = self.conv3d_layer(self.layer3, 64, 64, 3, 3, 3, [1, 1, 1, 1, 1], name="layer4", activation=activation, phase_train=phase_train) self.objectness = self.conv3D_to_output(self.layer4, 64, 2, 3, 3, 3, [1, 1, 1, 1, 1], name="objectness", activation=None) self.cordinate = self.conv3D_to_output(self.layer4, 64, 24, 3, 3, 3, [1, 1, 1, 1, 1], name="cordinate", activation=None) self.y = tf.nn.softmax(self.objectness, dim=-1) elif modelType == 1: # 3D-CNN(论文网络结构: 降采样1/4,即InputSize / OutputSize = 0.25) self.layer1 = self.conv3d_layer(voxel , 1, 10, 5, 5, 5, [1, 2, 2, 2, 1], name="layer1", activation=activation, phase_train=phase_train) self.layer2 = self.conv3d_layer(self.layer1, 10, 20, 5, 5, 5, [1, 2, 2, 2, 1], name="layer2", activation=activation, phase_train=phase_train) self.layer3 = self.conv3d_layer(self.layer2, 20, 30, 3, 3, 3, [1, 2, 2, 2, 1], name="layer3", activation=activation, phase_train=phase_train) base_shape = self.layer2.get_shape().as_list() obj_output_shape = [tf.shape(self.layer3)[0], base_shape[1], base_shape[2], base_shape[3], 2] cord_output_shape = [tf.shape(self.layer3)[0], base_shape[1], base_shape[2], base_shape[3], 24] self.objectness = self.deconv3D_to_output(self.layer3, 30, 2, 3, 3, 3, [1, 2, 2, 2, 1], obj_output_shape, name="objectness", activation=None) self.cordinate = self.deconv3D_to_output(self.layer3, 30, 24, 3, 3, 3, [1, 2, 2, 2, 1], cord_output_shape, name="cordinate", activation=None) self.y = tf.nn.softmax(self.objectness, dim=-1) # batch Normalize def batch_norm(self, inputs, phase_train = True, decay = 0.9, eps = 1e-5): ''' Inputs: inputs: input data for last layer phase_train: True / False, = True is train, = False is Test Outputs: norm data for next layer ''' gamma = tf.get_variable("gamma", shape=inputs.get_shape()[-1], dtype=tf.float32, initializer=tf.constant_initializer(1.0)) beta = tf.get_variable("beta", shape=inputs.get_shape()[-1], dtype=tf.float32, initializer=tf.constant_initializer(0.0)) pop_mean = tf.get_variable("pop_mean", trainable=False, shape=inputs.get_shape()[-1], dtype=tf.float32, initializer=tf.constant_initializer(0.0)) pop_var = tf.get_variable("pop_var", trainable=False, shape=inputs.get_shape()[-1], dtype=tf.float32, initializer=tf.constant_initializer(1.0)) axes = range(len(inputs.get_shape()) - 1) if phase_train == True: batch_mean, batch_var = tf.nn.moments(inputs, axes = [0, 1, 2, 3]) train_mean = tf.assign(pop_mean, pop_mean * decay + batch_mean*(1 - decay)) train_var = tf.assign(pop_var, pop_var * decay + batch_var * (1 - decay)) with tf.control_dependencies([train_mean, train_var]): return tf.nn.batch_normalization(inputs, batch_mean, batch_var, beta, gamma, eps) else: return tf.nn.batch_normalization(inputs, pop_mean, pop_var, beta, gamma, eps) # 3D Conv Layer def conv3d_layer(self, inputs, inputs_dims, outputs_dims, height, width, length, stride, activation=tf.nn.relu, padding="SAME", name="", phase_train = True): ''' Inputs: inputs: pre-Layer output inputs_dims: pre-Layer output channels outputs_dims: cur-Layer output channels [length, height, width]: cur-Layer conv3d kernel size stride: conv3d kernel move step in length/height/width axis activation: default use relu activation function padding: conv3d 'padding' parameter Outputs: 3D Conv. Layer outputs ''' with tf.variable_scope("conv3D" + name): # conv3d layer kernel kernel = tf.get_variable("weights", shape=[length, height, width, inputs_dims, outputs_dims], dtype = tf.float32, initializer=tf.truncated_normal_initializer(stddev=0.01)) # conv3d layer bias bias = tf.get_variable("bias", shape=[outputs_dims], dtype=tf.float32, initializer=tf.constant_initializer(0.0)) # conv3d conv = tf.nn.conv3d(inputs, kernel, stride, padding=padding) bias = tf.nn.bias_add(conv, bias) if activation: bias = activation(bias, name="activation") bias = self.batch_norm(bias, phase_train) return bias # 3D Conv to Classification Layer def conv3D_to_output(self, inputs, inputs_dims, outputs_dims, height, width, length, stride, activation=tf.nn.relu, padding="SAME", name="", phase_train = True): ''' Inputs: inputs: pre-Layer outputs inputs_dims: pre-Layer output channels outputs_dims: cur-Layer output channels stride: conv3d kernel move step in length/height/width axis activation: default use relu activation function padding: conv3d 'padding' parameter outputs_shape: de-conv outputs shape Outputs: conv outputs ''' with tf.variable_scope("conv3D" + name): kernel = tf.get_variable("weights", shape=[length, height, width, inputs_dims, outputs_dims], dtype=tf.float32, initializer=tf.constant_initializer(0.01)) conv = tf.nn.conv3d(inputs, kernel, stride, padding=padding) return conv # 3D Deconv. to Classification Layer def deconv3D_to_output(self, inputs, inputs_dims, outputs_dims, height, width, length, stride, output_shape, activation=tf.nn.relu, padding="SAME", name="", phase_train = True): ''' Inputs: inputs: pre-Layer outputs inputs_dims: pre-Layer output channels outputs_dims: cur-Layer output channels stride: conv3d kernel move step in length/height/width axis activation: default use relu activation function padding: conv3d 'padding' parameter outputs_shape: de-conv outputs shape Outputs: de-conv outputs ''' with tf.variable_scope("deconv3D"+name): kernel = tf.get_variable("weights", shape=[length, height, width, outputs_dims, inputs_dims], dtype=tf.float32, initializer=tf.constant_initializer(0.01)) deconv = tf.nn.conv3d_transpose(inputs, kernel, output_shape, stride, padding="SAME") return deconv # define loss def loss_Fun(self, lossType = 0, cord_loss_weight = 0.02): ''' Inputs: lossType: = for difference loss Type cord_loss_weight: 0.02 Outputs: ''' if lossType == 0: # print("g_obj") # print(self.cordinate.get_shape()) g_obj = tf.placeholder(tf.float32, self.cordinate.get_shape().as_list()[:4]) g_cord = tf.placeholder(tf.float32, self.cordinate.get_shape().as_list()) non_g_obj = tf.subtract(tf.ones_like(g_obj, dtype=tf.float32), g_obj ) elosion = 0.00001 y = self.y is_obj_loss = -tf.reduce_sum(tf.multiply(g_obj , tf.log(y[:,:,:,:,0] + elosion))) # object loss non_obj_loss = tf.reduce_sum(tf.multiply(non_g_obj, tf.log(y[:, :, :, :, 0] + elosion))) # non-object loss cross_entropy = tf.add(is_obj_loss, non_obj_loss) obj_loss = cross_entropy cord_diff = tf.multiply(g_obj , tf.reduce_sum(tf.square(tf.subtract(self.cordinate, g_cord)), 4)) # cord loss cord_loss = tf.multiply(tf.reduce_sum(cord_diff), cord_loss_weight) return tf.add(obj_loss, cord_loss), obj_loss, cord_loss, is_obj_loss, non_obj_loss, g_obj, g_cord, y # Create Optimizer def create_optimizer(self, all_loss, optType = "Adam", learnRate = 0.001): ''' Inputs: all_loss: graph all_loss lr: learn rate Outputs: optimizer ''' if optType == "Adam": opt = tf.train.AdamOptimizer(learnRate) optimizer = opt.minimize(all_loss) return optimizer
2.2 数据预处理
'''Prepase KITTI data for 3D Object detection Ref: 3D Fully Convolutional Network for Vehicle Detection in Point Cloud Author: Shiwen He Date: 28 April 2018 ''' import numpy as np from kitti_object import kitti_object as kittiReader import kitti_util # lidar data => 3D Grid Voxel # filter lidar data by camera FoV def filter_camera_fov(pc): ''' Inputs: pc: n x 3 Outputs: filter_pc: m x 3, m <= 3 Notices: FoV: from -45 degree to 45 degree ''' logic_fov = np.logical_and((pc[:, 1] < pc[:, 0] - 0.27), (-pc[:, 1] < pc[:, 0] - 0.27)) filter_pc = pc[logic_fov] return filter_pc # filter lidar data by detection range def filter_lidar_range(pc, limitX, limitY, limitZ): ''' Inputs: pc: n x 3, limitX, limitY, limitZ: 1 x 2 Outputs: filter_pc: m x 3, m <= n ''' logic_x = np.logical_and(pc[:, 0] >= limitX[0], pc[:, 0] < limitX[1]) logic_y = np.logical_and(pc[:, 1] >= limitY[0], pc[:, 1] < limitY[1]) logic_z = np.logical_and(pc[:, 2] >= limitZ[0], pc[:, 2] < limitZ[1]) logic_xyz = np.logical_and(logic_x, np.logical_and(logic_y, logic_z)) filter_pc = pc[:, :3][logic_xyz] return filter_pc # filter center + corners def filter_center_corners(centers, corners, boxsizes, limitX, limitY, limitZ): ''' Inputs: centers: n x 3 corners: n x 8 x 3 limitX, limitY, limitZ: 1 x 2 Outputs: filter_centers: m x 3, m <= n filter_corners: m x 3, m <= n ''' logic_x = np.logical_and(centers[:, 0] >= limitX[0], centers[:, 0] < limitX[1]) logic_y = np.logical_and(centers[:, 1] >= limitY[0], centers[:, 1] < limitY[1]) logic_z = np.logical_and(centers[:, 2] >= limitZ[0], centers[:, 2] < limitZ[1]) logic_xyz = np.logical_and(logic_x, np.logical_and(logic_y, logic_z)) filter_centers_1 = centers[logic_xyz, :] filter_corners_1 = corners[logic_xyz, :, :] filter_boxsizes_1 = boxsizes[logic_xyz, :] shape_centers = filter_centers_1.shape; filter_centers = np.zeros([shape_centers[0], 3]) filter_corners = np.zeros([shape_centers[0], 8, 3]); filter_boxsizes = np.zeros([shape_centers[0], 3]); idx = 0 for idx2 in range(shape_centers[0]): logic_x = np.logical_and(filter_corners_1[idx2, :, 0] >= limitX[0], filter_corners_1[idx2, :, 0] < limitX[1]) logic_y = np.logical_and(filter_corners_1[idx2, :, 1] >= limitY[0], filter_corners_1[idx2, :, 1] < limitY[1]) logic_z = np.logical_and(filter_corners_1[idx2, :, 2] >= limitZ[0], filter_corners_1[idx2, :, 2] < limitZ[1]) logic_xyz = np.logical_and(logic_x, np.logical_and(logic_y, logic_z)) if logic_xyz.all(): filter_centers[idx, :3] = filter_centers_1[idx2, :] filter_corners[idx, :8, :3] = filter_corners_1[idx2, :, :] filter_boxsizes[idx, :3] = filter_boxsizes_1[idx2, :] idx += 1 if idx > 0: return filter_centers[:idx, :], filter_corners[:idx, :, :], filter_boxsizes[:idx, :] else: return None, None, None def filter_label(object3Ds, objectType = 'Car'): ''' Inputs: object3Ds: objectType: Outputs: centers, corners, rotatey ''' idx = 0 data = np.zeros([50, 7]).astype(np.float32) for iter in object3Ds: if iter.type == "DontCare": continue; if iter.type == objectType: # position data[idx, 0] = iter.t[0] data[idx, 1] = iter.t[1] data[idx, 2] = iter.t[2] # size data[idx, 3] = iter.h data[idx, 4] = iter.w data[idx, 5] = iter.l # rotate data[idx, 6] = iter.ry idx += 1 if idx > 0: return data[:idx, :3], data[:idx, 3:6], data[:idx, 6] else: return None, None, None def proj_to_velo(calib_data): """ Inputs: calib_data: Outputs: project matrix: from camera cordination to velodyne cordination """ rect = calib_data.R0; # calib_data["R0_rect"].reshape(3, 3) velo_to_cam = calib_data.V2C; # calib_data["Tr_velo_to_cam"].reshape(3, 4) inv_rect = np.linalg.inv(rect) inv_velo_to_cam = np.linalg.pinv(velo_to_cam[:, :3]) return np.dot(inv_velo_to_cam, inv_rect) # corners_3d def compute_3d_corners(centers, sizes, rotates): ''' Inputs: centers: rotates: sizes: Outputs: corners_3d: n x 8 x 3 array in Lidar coord. ''' # print(centers) corners = [] for place, rotate, sz in zip(centers, rotates, sizes): x, y, z = place h, w, l = sz if l > 10: continue corner = np.array([ [x - l / 2., y - w / 2., z], [x + l / 2., y - w / 2., z], [x - l / 2., y + w / 2., z], [x - l / 2., y - w / 2., z + h], [x - l / 2., y + w / 2., z + h], [x + l / 2., y + w / 2., z], [x + l / 2., y - w / 2., z + h], [x + l / 2., y + w / 2., z + h], ]) corner -= np.array([x, y, z]) rotate_matrix = np.array([ [np.cos(rotate), -np.sin(rotate), 0], [np.sin(rotate), np.cos(rotate), 0], [0, 0, 1] ]) a = np.dot(corner, rotate_matrix.transpose()) a += np.array([x, y, z]) corners.append(a) corners_3d = np.array(corners) return corners_3d # lidar data to 3D Grid Voxel def lidar_to_binary_voxel(pc, resolution, limitX, limitY, limitZ): ''' Inputs: pc: n x 3, resolution: 1 x 3, limitX, limitY, limitZ: 1 x 2 Outputs: voxel: shape is inputSize ''' voxel_pc = np.zeros_like(pc).astype(np.int32) # Compute PointCloud Position in 3D Grid voxel_pc[:, 0] = ((pc[:, 0] - limitX[0]) / resolution[0]).astype(np.int32) voxel_pc[:, 1] = ((pc[:, 1] - limitY[0]) / resolution[1]).astype(np.int32) voxel_pc[:, 2] = ((pc[:, 2] - limitZ[0]) / resolution[2]).astype(np.int32) # 3D Grid voxel = np.zeros((int(round(limitX[1] - limitX[0]) / resolution[0]), int(round(limitY[1] - limitY[0]) / resolution[1]), \ int(round((limitZ[1] - limitZ[0]) / resolution[2])))) # 3D Grid Value voxel[voxel_pc[:, 0], voxel_pc[:, 1], voxel_pc[:, 2]] = 1 return voxel # label center to 3D Grid Voxel Center(sphere) def center_to_sphere(centers, boxsize, scale, resolution, limitX, limitY, limitZ): ''' Inputs: center: n x 3 boxsize: n x 3 scale: 1 x 1, = outputSize / inputSize resolution: 1 x 3 limitX, limitY, limitZ: 1 x 2 Outputs: spheres: m x 3, m <= n ''' # from 3D Box's bottom center => 3D center move_center = centers.copy(); print("centers") print(centers) print("boxsize") print(boxsize) move_center[:, 2] = centers[:, 2] + boxsize[:, 0] / 2; # compute Label Center PointCloud Position in 3D Grid spheres = np.zeros_like(move_center).astype(np.int32) spheres[:, 0] = ((move_center[:, 0] - limitX[0]) / resolution[0] * scale).astype(np.int32) spheres[:, 1] = ((move_center[:, 1] - limitY[0]) / resolution[1] * scale).astype(np.int32) spheres[:, 2] = ((move_center[:, 2] - limitZ[0]) / resolution[2] * scale).astype(np.int32) print("move_center") print(move_center) print("spheres") print(spheres) return spheres # 3D Grid Voxel Center(sphere) to label center def sphere_to_center(spheres, scale, resolution, limitX, limitY, limitZ): ''' Inputs: spheres: n x 3 scale: 1 x 1, = outputSize / inputSize resolution: 1 x 3 limitX, limitY, limitZ: 1 x 2 Outputs: centers: m x 3, m <= 3 ''' centers = np.zeros_like(spheres).astype(np.float32); centers[:, 0] = spheres[:, 0] * resolution[0] / scale + limitX[0] centers[:, 1] = spheres[:, 1] * resolution[1] / scale + limitY[0] centers[:, 2] = spheres[:, 2] * resolution[2] / scale + limitZ[0] return centers # label corners to 3D Grid Voxel: corners - centers def corners_to_train(spheres, corners, scale, resolution, limitX, limitY, limitZ): ''' Inputs: spheres: n x 3 corners: n x 8 x 3 scale: 1 x 1, = outputSize / inputSize resolution: 1 x 3 limitX, limitY, limitZ: 1 x 2 Outputs: train_corners: m x 3, m <= n ''' # 3D Grid Voxel Center => label center centers = sphere_to_center(spheres, scale, resolution, limitX, limitY, limitZ) train_corners = np.zeros_like(corners).astype(np.float32) # train corners for regression loss for index, (corner, center) in enumerate(zip(corners, centers)): train_corners[index] = corner - center return train_corners # create center and cordination for train def create_train_label(centers, corners, boxsize, scale, resolution, limitX, limitY, limitZ): ''' Inputs: centers: n x 3 corners: n x 8 x 3 boxsize: n x 3 scale: 1 x 1, outputSize / inputSize resolution: 1 x 3 limitX. limitY, limitZ: 1 x 2 Outputs: train_centers: m x 3, m <= n train_corners: m x 3, m <= n ''' train_centers = center_to_sphere(centers, boxsize, scale, resolution, limitX, limitY, limitZ) train_corners = corners_to_train(train_centers, corners, scale, resolution, limitX, limitY, limitZ) return train_centers, train_corners def create_obj_map(train_centers, scale, resolution, limitX, limitY, limitZ): ''' Inputs: centers: n x 3 scale: 1 x 1, outputSize / inputSize resolution: 1 x 3 limitX, limitY, limitZ: 1 x 2 Outputs: obj_map: shape is scale * inputSize ''' # 3D Grid sizeX = int(round((limitX[1] - limitX[0]) / resolution[0] * scale)) sizeY = int(round((limitY[1] - limitY[0]) / resolution[1] * scale)) sizeZ = int(round((limitZ[1] - limitZ[0]) / resolution[2] * scale)) obj_map = np.zeros([sizeX, sizeY, sizeZ]) # print("sizeX, sizeY, sizeZ") # print(sizeX, sizeY, sizeZ) # objectness map: label center in objectness map where value is 1 obj_map[train_centers[:,0], train_centers[:, 1], train_centers[:, 2]] = 1; return obj_map def create_cord_map(train_centers, train_corners, scale, resolution, limitX, limitY, limitZ): ''' Inputs: train_centers: n x 3 train_corners: n x 8 x 3 scale: 1 x 1, outputSize / inputSize resolution: 1 x 3 limitX, limitY, limitZ: 1 x 2 Outputs: cord_map: shape is inputSize * scale ''' # reshape train_corners: n x 8 x 3 => n x 24 corners = train_corners.reshape(train_corners.shape[0], -1) # 3D Grid sizeX = int(round((limitX[1] - limitX[0]) / resolution[0] * scale)) sizeY = int(round((limitY[1] - limitY[0]) / resolution[1] * scale)) sizeZ = int(round((limitZ[1] - limitZ[0]) / resolution[2] * scale)) sizeD = 24 cord_map = np.zeros([sizeX, sizeY, sizeZ, sizeD]) # print(train_centers) cord_map[train_centers[:,0], train_centers[:, 1], train_centers[:, 2]] = corners return cord_map # kitti data interface: class kitti_3DVoxel_interface(object): def __init__(self, root_dir, objectType = 'Car', split='training', scale = 0.25, resolution = (0.2, 0.2, 0.2), limitX = (0, 80), limitY = (-40, 40), limitZ = (-2.5, 1.5)): ''' Inputs: case1 root_dir: train or val. data dir, train or val.'s file struct like: root_dir->training->velodyne root_dir->training->calib root_dir->training->label_2 case2 root_dir: test data dir, test's file struct like: root_dir->testing->velodyne root_dir->testing->calib Outputs: -None ''' self.root_dir = root_dir self.split = split self.object = kittiReader(self.root_dir, self.split) self.objectType = objectType self.scale = scale self.resolution = resolution self.limitX = limitX self.limitY = limitY self.limitZ = limitZ def read_kitti_data(self, idx = 0): ''' Inputs: idx: training or testing sample index Outputs: voxel : inputSize obj_map : scale * inputSize cord_map : scale * inputSize ''' kitti_Object3Ds = None kitti_Lidar = None kitti_Calib = None if self.split == 'training': # read Lidar data + Lidar Label + Calib data kitti_Object3Ds = self.object.get_label_objects(idx); kitti_Lidar = self.object.get_lidar(idx); kitti_Calib = self.object.get_calibration(idx); # lidar data filter filter_fov = filter_camera_fov(kitti_Lidar) filter_range = filter_lidar_range(filter_fov, self.limitX, self.limitY, self.limitZ) # label filter centers, boxsizes, rotates = filter_label(kitti_Object3Ds, self.objectType) if centers is None: return None, None, None # label center: Notice from camera Coordination to velo. Coordination if not(kitti_Calib is None): proj_velo = proj_to_velo(kitti_Calib)[:, :3] centers = np.dot(centers, proj_velo.transpose())[:, :3] # label corners: corners = compute_3d_corners(centers, boxsizes, rotates) # print(corners) # print(corners.shape) # filter centers + corners filter_centers, filter_corners, boxsizes = filter_center_corners(centers, corners, boxsizes, self.limitX, self.limitY, self.limitZ) # print(filter_centers) # print(filter_corners) if not(filter_centers is None): # training center train_centers, train_corners = create_train_label(filter_centers, filter_corners, boxsizes, self.scale, self.resolution, self.limitX, self.limitY, self.limitZ) # print("filter_centers") # print(filter_centers) # print("train_centers") # print(train_centers) # obj_map / cord_map / voxel obj_map = create_obj_map(train_centers, self.scale, self.resolution, self.limitX, self.limitY, self.limitZ) cord_map = create_cord_map(train_centers, train_corners, self.scale, self.resolution, self.limitX, self.limitY, self.limitZ) voxel = lidar_to_binary_voxel(filter_range, self.resolution, self.limitX, self.limitY, self.limitZ) return voxel, obj_map, cord_map else: return None, None, None elif self.split == 'testing': # read Lidar Data + Calib + Data kitti_Lidar = self.object.get_lidar(idx); kitti_Calib = self.object.get_calibration(idx); # lidar data filter filter_fov = filter_camera_fov(kitti_Lidar) filter_range = filter_lidar_range(filter_fov, self.limitX, self.limitY, self.limitZ) voxel = lidar_to_binary_voxel(filter_range, self.resolution, self.limitX, self.limitY, self.limitZ) return voxel if __name__ == '__main__': data_dir = "/home/hsw/桌面/PCL_API_Doc/frustum-pointnets-master/dataset" kitti_3DVoxel = kitti_3DVoxel_interface(data_dir, objectType = 'Car', split='training', scale = 0.25, resolution = (0.2, 0.2, 0.2), limitX = (0, 80), limitY = (-40, 40), limitZ = (-2.5, 1.5)) sampleIdx = 195; voxel, obj_map, cord_map = kitti_3DVoxel.read_kitti_data(sampleIdx) if not(voxel is None): print(voxel.shape) print(obj_map.shape) print(cord_map.shape)
2.3 KITTI数据读取相关
''' Helper class and functions for loading KITTI objects Author: Charles R. Qi Date: September 2017 ''' from __future__ import print_function import os import sys import numpy as np import cv2 from PIL import Image BASE_DIR = os.path.dirname(os.path.abspath(__file__)) ROOT_DIR = os.path.dirname(BASE_DIR) sys.path.append(os.path.join(ROOT_DIR, 'mayavi')) import kitti_util as utils try: raw_input # Python 2 except NameError: raw_input = input # Python 3 # 3D static data class kitti_object(object): '''Load and parse object data into a usable format.''' def __init__(self, root_dir, split='training'): '''root_dir contains training and testing folders''' self.root_dir = root_dir self.split = split self.split_dir = os.path.join(root_dir, split) if split == 'training': self.num_samples = 7481 elif split == 'testing': self.num_samples = 7518 else: print('Unknown split: %s' % (split)) exit(-1) # data dir self.image_dir = os.path.join(self.split_dir, 'image_2') self.calib_dir = os.path.join(self.split_dir, 'calib') self.lidar_dir = os.path.join(self.split_dir, 'velodyne') self.label_dir = os.path.join(self.split_dir, 'label_2') def __len__(self): return self.num_samples # read image: return image def get_image(self, idx): assert(idx<self.num_samples) img_filename = os.path.join(self.image_dir, '%06d.png'%(idx)) return utils.load_image(img_filename) # read lidar: return n x 4 def get_lidar(self, idx): assert(idx<self.num_samples) lidar_filename = os.path.join(self.lidar_dir, '%06d.bin'%(idx)) return utils.load_velo_scan(lidar_filename) # read calib file: def get_calibration(self, idx): assert(idx<self.num_samples) calib_filename = os.path.join(self.calib_dir, '%06d.txt'%(idx)) return utils.Calibration(calib_filename) # read label def get_label_objects(self, idx): assert(idx<self.num_samples and self.split=='training') label_filename = os.path.join(self.label_dir, '%06d.txt'%(idx)) return utils.read_label(label_filename) # read depth map def get_depth_map(self, idx): pass # read top_down image def get_top_down(self, idx): pass class kitti_object_video(object): ''' Load data for KITTI videos ''' def __init__(self, img_dir, lidar_dir, calib_dir): self.calib = utils.Calibration(calib_dir, from_video=True) self.img_dir = img_dir self.lidar_dir = lidar_dir self.img_filenames = sorted([os.path.join(img_dir, filename) \ for filename in os.listdir(img_dir)]) self.lidar_filenames = sorted([os.path.join(lidar_dir, filename) \ for filename in os.listdir(lidar_dir)]) print(len(self.img_filenames)) print(len(self.lidar_filenames)) #assert(len(self.img_filenames) == len(self.lidar_filenames)) self.num_samples = len(self.img_filenames) def __len__(self): return self.num_samples def get_image(self, idx): assert(idx<self.num_samples) img_filename = self.img_filenames[idx] return utils.load_image(img_filename) def get_lidar(self, idx): assert(idx<self.num_samples) lidar_filename = self.lidar_filenames[idx] return utils.load_velo_scan(lidar_filename) def get_calibration(self, unused): return self.calib def viz_kitti_video(): video_path = os.path.join(ROOT_DIR, 'dataset/2011_09_26/') dataset = kitti_object_video(\ os.path.join(video_path, '2011_09_26_drive_0023_sync/image_02/data'), os.path.join(video_path, '2011_09_26_drive_0023_sync/velodyne_points/data'), video_path) print(len(dataset)) for i in range(len(dataset)): img = dataset.get_image(0) pc = dataset.get_lidar(0) Image.fromarray(img).show() draw_lidar(pc) raw_input() pc[:,0:3] = dataset.get_calibration().project_velo_to_rect(pc[:,0:3]) draw_lidar(pc) raw_input() return def show_image_with_boxes(img, objects, calib, show3d=True): ''' Show image with 2D bounding boxes ''' img1 = np.copy(img) # for 2d bbox img2 = np.copy(img) # for 3d bbox for obj in objects: if obj.type=='DontCare':continue cv2.rectangle(img1, (int(obj.xmin),int(obj.ymin)), (int(obj.xmax),int(obj.ymax)), (0,255,0), 2) box3d_pts_2d, box3d_pts_3d = utils.compute_box_3d(obj, calib.P) img2 = utils.draw_projected_box3d(img2, box3d_pts_2d) Image.fromarray(img1).show() if show3d: Image.fromarray(img2).show() def get_lidar_in_image_fov(pc_velo, calib, xmin, ymin, xmax, ymax, return_more=False, clip_distance=2.0): ''' Filter lidar points, keep those in image FOV ''' pts_2d = calib.project_velo_to_image(pc_velo) fov_inds = (pts_2d[:,0]<xmax) & (pts_2d[:,0]>=xmin) & \ (pts_2d[:,1]<ymax) & (pts_2d[:,1]>=ymin) fov_inds = fov_inds & (pc_velo[:,0]>clip_distance) imgfov_pc_velo = pc_velo[fov_inds,:] if return_more: return imgfov_pc_velo, pts_2d, fov_inds else: return imgfov_pc_velo def show_lidar_with_boxes(pc_velo, objects, calib, img_fov=False, img_width=None, img_height=None): ''' Show all LiDAR points. Draw 3d box in LiDAR point cloud (in velo coord system) ''' if 'mlab' not in sys.modules: import mayavi.mlab as mlab from viz_util import draw_lidar_simple, draw_lidar, draw_gt_boxes3d print(('All point num: ', pc_velo.shape[0])) fig = mlab.figure(figure=None, bgcolor=(0,0,0), fgcolor=None, engine=None, size=(1000, 500)) if img_fov: pc_velo = get_lidar_in_image_fov(pc_velo, calib, 0, 0, img_width, img_height) print(('FOV point num: ', pc_velo.shape[0])) draw_lidar(pc_velo, fig=fig) for obj in objects: if obj.type=='DontCare':continue # Draw 3d bounding box box3d_pts_2d, box3d_pts_3d = utils.compute_box_3d(obj, calib.P) box3d_pts_3d_velo = calib.project_rect_to_velo(box3d_pts_3d) # Draw heading arrow ori3d_pts_2d, ori3d_pts_3d = utils.compute_orientation_3d(obj, calib.P) ori3d_pts_3d_velo = calib.project_rect_to_velo(ori3d_pts_3d) x1,y1,z1 = ori3d_pts_3d_velo[0,:] x2,y2,z2 = ori3d_pts_3d_velo[1,:] draw_gt_boxes3d([box3d_pts_3d_velo], fig=fig) mlab.plot3d([x1, x2], [y1, y2], [z1,z2], color=(0.5,0.5,0.5), tube_radius=None, line_width=1, figure=fig) mlab.show(1) def show_lidar_on_image(pc_velo, img, calib, img_width, img_height): ''' Project LiDAR points to image ''' imgfov_pc_velo, pts_2d, fov_inds = get_lidar_in_image_fov(pc_velo, calib, 0, 0, img_width, img_height, True) imgfov_pts_2d = pts_2d[fov_inds,:] imgfov_pc_rect = calib.project_velo_to_rect(imgfov_pc_velo) import matplotlib.pyplot as plt cmap = plt.cm.get_cmap('hsv', 256) cmap = np.array([cmap(i) for i in range(256)])[:,:3]*255 for i in range(imgfov_pts_2d.shape[0]): depth = imgfov_pc_rect[i,2] color = cmap[int(640.0/depth),:] cv2.circle(img, (int(np.round(imgfov_pts_2d[i,0])), int(np.round(imgfov_pts_2d[i,1]))), 2, color=tuple(color), thickness=-1) Image.fromarray(img).show() return img def dataset_viz(): dataset = kitti_object(os.path.join(ROOT_DIR, 'dataset/KITTI/object')) for data_idx in range(len(dataset)): # Load data from dataset objects = dataset.get_label_objects(data_idx) objects[0].print_object() img = dataset.get_image(data_idx) img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) img_height, img_width, img_channel = img.shape print(('Image shape: ', img.shape)) pc_velo = dataset.get_lidar(data_idx)[:,0:3] calib = dataset.get_calibration(data_idx) # Draw 2d and 3d boxes on image show_image_with_boxes(img, objects, calib, False) raw_input() # Show all LiDAR points. Draw 3d box in LiDAR point cloud show_lidar_with_boxes(pc_velo, objects, calib, True, img_width, img_height) raw_input() if __name__=='__main__': import mayavi.mlab as mlab from viz_util import draw_lidar_simple, draw_lidar, draw_gt_boxes3d dataset_viz()
""" Helper methods for loading and parsing KITTI data. Author: Charles R. Qi Date: September 2017 """ from __future__ import print_function import numpy as np import cv2 import os class Object3d(object): ''' 3d object label ''' def __init__(self, label_file_line): data = label_file_line.split(' ') data[1:] = [float(x) for x in data[1:]] # extract label, truncation, occlusion self.type = data[0] # 'Car', 'Pedestrian', ... self.truncation = data[1] # truncated pixel ratio [0..1] self.occlusion = int(data[2]) # 0=visible, 1=partly occluded, 2=fully occluded, 3=unknown self.alpha = data[3] # object observation angle [-pi..pi] # extract 2d bounding box in 0-based coordinates self.xmin = data[4] # left self.ymin = data[5] # top self.xmax = data[6] # right self.ymax = data[7] # bottom self.box2d = np.array([self.xmin,self.ymin,self.xmax,self.ymax]) # extract 3d bounding box information self.h = data[8] # box height self.w = data[9] # box width self.l = data[10] # box length (in meters) self.t = (data[11],data[12],data[13]) # location (x,y,z) in camera coord. self.ry = data[14] # yaw angle (around Y-axis in camera coordinates) [-pi..pi] def print_object(self): print('Type, truncation, occlusion, alpha: %s, %d, %d, %f' % \ (self.type, self.truncation, self.occlusion, self.alpha)) print('2d bbox (x0,y0,x1,y1): %f, %f, %f, %f' % \ (self.xmin, self.ymin, self.xmax, self.ymax)) print('3d bbox h,w,l: %f, %f, %f' % \ (self.h, self.w, self.l)) print('3d bbox location, ry: (%f, %f, %f), %f' % \ (self.t[0],self.t[1],self.t[2],self.ry)) class Calibration(object): ''' Calibration matrices and utils 3d XYZ in <label>.txt are in rect camera coord. 2d box xy are in image2 coord Points in <lidar>.bin are in Velodyne coord. y_image2 = P^2_rect * x_rect y_image2 = P^2_rect * R0_rect * Tr_velo_to_cam * x_velo x_ref = Tr_velo_to_cam * x_velo x_rect = R0_rect * x_ref P^2_rect = [f^2_u, 0, c^2_u, -f^2_u b^2_x; 0, f^2_v, c^2_v, -f^2_v b^2_y; 0, 0, 1, 0] = K * [1|t] image2 coord: ----> x-axis (u) | | v y-axis (v) velodyne coord: front x, left y, up z rect/ref camera coord: right x, down y, front z Ref (KITTI paper): http://www.cvlibs.net/publications/Geiger2013IJRR.pdf TODO(rqi): do matrix multiplication only once for each projection. ''' def __init__(self, calib_filepath, from_video=False): if from_video: calibs = self.read_calib_from_video(calib_filepath) else: calibs = self.read_calib_file(calib_filepath) # Projection matrix from rect camera coord to image2 coord self.P = calibs['P2'] self.P = np.reshape(self.P, [3,4]) # Rigid transform from Velodyne coord to reference camera coord self.V2C = calibs['Tr_velo_to_cam'] self.V2C = np.reshape(self.V2C, [3,4]) self.C2V = inverse_rigid_trans(self.V2C) # Rotation from reference camera coord to rect camera coord self.R0 = calibs['R0_rect'] self.R0 = np.reshape(self.R0,[3,3]) # Camera intrinsics and extrinsics self.c_u = self.P[0,2] self.c_v = self.P[1,2] self.f_u = self.P[0,0] self.f_v = self.P[1,1] self.b_x = self.P[0,3]/(-self.f_u) # relative self.b_y = self.P[1,3]/(-self.f_v) def read_calib_file(self, filepath): ''' Read in a calibration file and parse into a dictionary. Ref: https://github.com/utiasSTARS/pykitti/blob/master/pykitti/utils.py ''' data = {} with open(filepath, 'r') as f: for line in f.readlines(): line = line.rstrip() if len(line)==0: continue key, value = line.split(':', 1) # The only non-float values in these files are dates, which # we don't care about anyway try: data[key] = np.array([float(x) for x in value.split()]) except ValueError: pass return data def read_calib_from_video(self, calib_root_dir): ''' Read calibration for camera 2 from video calib files. there are calib_cam_to_cam and calib_velo_to_cam under the calib_root_dir ''' data = {} cam2cam = self.read_calib_file(os.path.join(calib_root_dir, 'calib_cam_to_cam.txt')) velo2cam = self.read_calib_file(os.path.join(calib_root_dir, 'calib_velo_to_cam.txt')) Tr_velo_to_cam = np.zeros((3,4)) Tr_velo_to_cam[0:3,0:3] = np.reshape(velo2cam['R'], [3,3]) Tr_velo_to_cam[:,3] = velo2cam['T'] data['Tr_velo_to_cam'] = np.reshape(Tr_velo_to_cam, [12]) data['R0_rect'] = cam2cam['R_rect_00'] data['P2'] = cam2cam['P_rect_02'] return data def cart2hom(self, pts_3d): ''' Input: nx3 points in Cartesian Oupput: nx4 points in Homogeneous by pending 1 ''' n = pts_3d.shape[0] pts_3d_hom = np.hstack((pts_3d, np.ones((n,1)))) return pts_3d_hom # =========================== # ------- 3d to 3d ---------- # =========================== def project_velo_to_ref(self, pts_3d_velo): pts_3d_velo = self.cart2hom(pts_3d_velo) # nx4 return np.dot(pts_3d_velo, np.transpose(self.V2C)) def project_ref_to_velo(self, pts_3d_ref): pts_3d_ref = self.cart2hom(pts_3d_ref) # nx4 return np.dot(pts_3d_ref, np.transpose(self.C2V)) def project_rect_to_ref(self, pts_3d_rect): ''' Input and Output are nx3 points ''' return np.transpose(np.dot(np.linalg.inv(self.R0), np.transpose(pts_3d_rect))) def project_ref_to_rect(self, pts_3d_ref): ''' Input and Output are nx3 points ''' return np.transpose(np.dot(self.R0, np.transpose(pts_3d_ref))) def project_rect_to_velo(self, pts_3d_rect): ''' Input: nx3 points in rect camera coord. Output: nx3 points in velodyne coord. ''' pts_3d_ref = self.project_rect_to_ref(pts_3d_rect) return self.project_ref_to_velo(pts_3d_ref) def project_velo_to_rect(self, pts_3d_velo): pts_3d_ref = self.project_velo_to_ref(pts_3d_velo) return self.project_ref_to_rect(pts_3d_ref) # =========================== # ------- 3d to 2d ---------- # =========================== def project_rect_to_image(self, pts_3d_rect): ''' Input: nx3 points in rect camera coord. Output: nx2 points in image2 coord. ''' pts_3d_rect = self.cart2hom(pts_3d_rect) pts_2d = np.dot(pts_3d_rect, np.transpose(self.P)) # nx3 pts_2d[:,0] /= pts_2d[:,2] pts_2d[:,1] /= pts_2d[:,2] return pts_2d[:,0:2] def project_velo_to_image(self, pts_3d_velo): ''' Input: nx3 points in velodyne coord. Output: nx2 points in image2 coord. ''' pts_3d_rect = self.project_velo_to_rect(pts_3d_velo) return self.project_rect_to_image(pts_3d_rect) # =========================== # ------- 2d to 3d ---------- # =========================== def project_image_to_rect(self, uv_depth): ''' Input: nx3 first two channels are uv, 3rd channel is depth in rect camera coord. Output: nx3 points in rect camera coord. ''' n = uv_depth.shape[0] x = ((uv_depth[:,0]-self.c_u)*uv_depth[:,2])/self.f_u + self.b_x y = ((uv_depth[:,1]-self.c_v)*uv_depth[:,2])/self.f_v + self.b_y pts_3d_rect = np.zeros((n,3)) pts_3d_rect[:,0] = x pts_3d_rect[:,1] = y pts_3d_rect[:,2] = uv_depth[:,2] return pts_3d_rect def project_image_to_velo(self, uv_depth): pts_3d_rect = self.project_image_to_rect(uv_depth) return self.project_rect_to_velo(pts_3d_rect) def rotx(t): ''' 3D Rotation about the x-axis. ''' c = np.cos(t) s = np.sin(t) return np.array([[1, 0, 0], [0, c, -s], [0, s, c]]) def roty(t): ''' Rotation about the y-axis. ''' c = np.cos(t) s = np.sin(t) return np.array([[c, 0, s], [0, 1, 0], [-s, 0, c]]) def rotz(t): ''' Rotation about the z-axis. ''' c = np.cos(t) s = np.sin(t) return np.array([[c, -s, 0], [s, c, 0], [0, 0, 1]]) def transform_from_rot_trans(R, t): ''' Transforation matrix from rotation matrix and translation vector. ''' R = R.reshape(3, 3) t = t.reshape(3, 1) return np.vstack((np.hstack([R, t]), [0, 0, 0, 1])) def inverse_rigid_trans(Tr): ''' Inverse a rigid body transform matrix (3x4 as [R|t]) [R'|-R't; 0|1] ''' inv_Tr = np.zeros_like(Tr) # 3x4 inv_Tr[0:3,0:3] = np.transpose(Tr[0:3,0:3]) inv_Tr[0:3,3] = np.dot(-np.transpose(Tr[0:3,0:3]), Tr[0:3,3]) return inv_Tr def read_label(label_filename): lines = [line.rstrip() for line in open(label_filename)] objects = [Object3d(line) for line in lines] return objects def load_image(img_filename): return cv2.imread(img_filename) def load_velo_scan(velo_filename): scan = np.fromfile(velo_filename, dtype=np.float32) scan = scan.reshape((-1, 4)) return scan def project_to_image(pts_3d, P): ''' Project 3d points to image plane. Usage: pts_2d = projectToImage(pts_3d, P) input: pts_3d: nx3 matrix P: 3x4 projection matrix output: pts_2d: nx2 matrix P(3x4) dot pts_3d_extended(4xn) = projected_pts_2d(3xn) => normalize projected_pts_2d(2xn) <=> pts_3d_extended(nx4) dot P'(4x3) = projected_pts_2d(nx3) => normalize projected_pts_2d(nx2) ''' n = pts_3d.shape[0] pts_3d_extend = np.hstack((pts_3d, np.ones((n,1)))) print(('pts_3d_extend shape: ', pts_3d_extend.shape)) pts_2d = np.dot(pts_3d_extend, np.transpose(P)) # nx3 pts_2d[:,0] /= pts_2d[:,2] pts_2d[:,1] /= pts_2d[:,2] return pts_2d[:,0:2] # corners_2d + corners_3d def compute_box_3d(obj, P): ''' Takes an object and a projection matrix (P) and projects the 3d bounding box into the image plane. Returns: corners_2d: (8,2) array in left image coord. corners_3d: (8,3) array in in rect camera coord. ''' # compute rotational matrix around yaw axis R = roty(obj.ry) # 3d bounding box dimensions l = obj.l; w = obj.w; h = obj.h; # 3d bounding box corners x_corners = [l/2,l/2,-l/2,-l/2,l/2,l/2,-l/2,-l/2]; y_corners = [0,0,0,0,-h,-h,-h,-h]; z_corners = [w/2,-w/2,-w/2,w/2,w/2,-w/2,-w/2,w/2]; # rotate and translate 3d bounding box corners_3d = np.dot(R, np.vstack([x_corners,y_corners,z_corners])) #print corners_3d.shape corners_3d[0,:] = corners_3d[0,:] + obj.t[0]; corners_3d[1,:] = corners_3d[1,:] + obj.t[1]; corners_3d[2,:] = corners_3d[2,:] + obj.t[2]; #print 'cornsers_3d: ', corners_3d # only draw 3d bounding box for objs in front of the camera if np.any(corners_3d[2,:]<0.1): corners_2d = None return corners_2d, np.transpose(corners_3d) # project the 3d bounding box into the image plane corners_2d = project_to_image(np.transpose(corners_3d), P); #print 'corners_2d: ', corners_2d return corners_2d, np.transpose(corners_3d) def compute_orientation_3d(obj, P): ''' Takes an object and a projection matrix (P) and projects the 3d object orientation vector into the image plane. Returns: orientation_2d: (2,2) array in left image coord. orientation_3d: (2,3) array in in rect camera coord. ''' # compute rotational matrix around yaw axis R = roty(obj.ry) # orientation in object coordinate system orientation_3d = np.array([[0.0, obj.l],[0,0],[0,0]]) # rotate and translate in camera coordinate system, project in image orientation_3d = np.dot(R, orientation_3d) orientation_3d[0,:] = orientation_3d[0,:] + obj.t[0] orientation_3d[1,:] = orientation_3d[1,:] + obj.t[1] orientation_3d[2,:] = orientation_3d[2,:] + obj.t[2] # vector behind image plane? if np.any(orientation_3d[2,:]<0.1): orientation_2d = None return orientation_2d, np.transpose(orientation_3d) # project orientation into the image plane orientation_2d = project_to_image(np.transpose(orientation_3d), P); return orientation_2d, np.transpose(orientation_3d) def draw_projected_box3d(image, qs, color=(255,255,255), thickness=2): ''' Draw 3d bounding box in image qs: (8,3) array of vertices for the 3d box in following order: 1 -------- 0 /| /| 2 -------- 3 . | | | | . 5 -------- 4 |/ |/ 6 -------- 7 ''' qs = qs.astype(np.int32) for k in range(0,4): # Ref: http://docs.enthought.com/mayavi/mayavi/auto/mlab_helper_functions.html i,j=k,(k+1)%4 # use LINE_AA for opencv3 cv2.line(image, (qs[i,0],qs[i,1]), (qs[j,0],qs[j,1]), color, thickness, cv2.CV_AA) i,j=k+4,(k+1)%4 + 4 cv2.line(image, (qs[i,0],qs[i,1]), (qs[j,0],qs[j,1]), color, thickness, cv2.CV_AA) i,j=k,k+4 cv2.line(image, (qs[i,0],qs[i,1]), (qs[j,0],qs[j,1]), color, thickness, cv2.CV_AA) return image
3. 通过测试还在训练,但是我的硬件设备较差,所以,训练速度比较慢