本文分享在DAIR-V2X-V數(shù)據(jù)集中，將標(biāo)簽轉(zhuǎn)為Kitti格式，并可視化3D檢測效果。
一、將標(biāo)簽轉(zhuǎn)為Kitti格式
DAIR-V2X包括不同類型的數(shù)據(jù)集：
- DAIR-V2X-I
- DAIR-V2X-V
- DAIR-V2X-C
- V2X-Seq-SPD
- V2X-Seq-TFD
- DAIR-V2X-C-Example:?google_drive_link
- V2X-Seq-SPD-Example:?google_drive_link
- V2X-Seq-TFD-Example:?google_drive_link
本文選擇DAIR-V2X-V作為示例。
1、下載DAIR-V2X工程
?DAIR-V2X開源地址：https://github.com/AIR-THU/DAIR-V2X
2、存放數(shù)據(jù)
可以將數(shù)據(jù)存放到data目錄中，比如：data/DAIR-V2X-V/single-vehicle-side，里面包含兩個關(guān)鍵目錄和一個文件
calib/
label/
data_info.json
3、修復(fù)bug
在tools/dataset_converter/gen_kitti/label_json2kitti.py中的22行，將 i15 = str(-eval(item["rotation"])) 改為：
i15 = str(-float(item["rotation"]))如何不修改會報錯的；
DAIR-V2X-gp/tools/dataset_converter/gen_kitti/label_json2kitti.py", line 22, in write_kitti_in_txt
? ? i15 = str(-eval(item["rotation"]))
TypeError: eval() arg 1 must be a string, bytes or code object
將tools/dataset_converter/gen_kitti/label_json2kitti.py復(fù)制到根目錄中，避免找不到tool庫。
4、修改配置參數(shù)
label_json2kitti.py中，可以將rawdata_copy和kitti_pcd2bin注釋掉。
這樣節(jié)約時間，不用程序拷貝圖像、點云數(shù)據(jù)，只需生成標(biāo)簽即可。
if __name__ == "__main__":
print("================ Start to Convert ================")
args = parser.parse_args()
source_root = args.source_root
target_root = args.target_root
print("================ Start to Copy Raw Data ================")
mdkir_kitti(target_root)
# rawdata_copy(source_root, target_root)
# kitti_pcd2bin(target_root)
5、轉(zhuǎn)換數(shù)據(jù)
執(zhí)行如下命令
python dair2kitti.py --source-root ./data/DAIR-V2X-V/single-vehicle-side --target-root ./data/DAIR-V2X-V/single-vehicle-side --split-path ./data/split_datas/single-vehicle-split-data.json --label-type camera --sensor-view vehicle
會打印信息
================ Start to Convert ================
================ Start to Copy Raw Data ================
================ Start to Generate Label ================
================ Start to Generate Calibration Files ================
15627 15627
================ Start to Generate ImageSet Files ================
查看目錄：data/DAIR-V2X-V/single-vehicle-side，生成了3個目錄
ImageSets
testing
training
其中，testing目錄是空的
ImageSets目錄包含：
training目錄包含：
6、查看生成數(shù)據(jù)格式
查看calib中的相機標(biāo)定文件，比如 000000.txt
P2: 3996.487567 0.0 955.58618 0.0 0.0 3963.430994 527.646219 0.0 0.0 0.0 1.0 0.0
P2: 3996.487567 0.0 955.58618 0.0 0.0 3963.430994 527.646219 0.0 0.0 0.0 1.0 0.0
P2: 3996.487567 0.0 955.58618 0.0 0.0 3963.430994 527.646219 0.0 0.0 0.0 1.0 0.0
P2: 3996.487567 0.0 955.58618 0.0 0.0 3963.430994 527.646219 0.0 0.0 0.0 1.0 0.0
R0_rect: 1 0 0 0 1 0 0 0 1
Tr_velo_to_cam: 0.006283 -0.999979 -0.001899 -0.298036 -0.005334 0.001865 -0.999984 -0.666812 0.999966 0.006293 -0.005322 -0.516927
Tr_velo_to_cam: 0.006283 -0.999979 -0.001899 -0.298036 -0.005334 0.001865 -0.999984 -0.666812 0.999966 0.006293 -0.005322 -0.516927
查看lable_2中的標(biāo)簽，比如 000000.txt
Car 0 0 0.33888581543844903 0 527.938232 69.723068 637.4556269999999 0.850836 4.337498 2.073565 -9.601712831407 0.8624079931420001 32.383280568744 1.615145
二、可視化3D框
?使用Kitti的方式，實現(xiàn)可視化推理結(jié)果，上面生成的結(jié)果，和kitii標(biāo)簽格式是一致的。
在新建一個vis目錄包括：
dataset? ? ? ? ? ? ? ? ? ? 存放相機標(biāo)定數(shù)據(jù)、圖片、標(biāo)簽
save_3d_output? 存放可視化圖片
kitti_3d_vis.py? ? ?可視化運行此代碼
kitti_util.py? ? ? ? ? ? 依賴代碼
具體的目錄結(jié)構(gòu)：
主代碼 kitti_3d_vis.py
# kitti_3d_vis.py
from __future__ import print_function
import os
import sys
import cv2
import random
import os.path
import shutil
from PIL import Image
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
ROOT_DIR = os.path.dirname(BASE_DIR)
sys.path.append(BASE_DIR)
sys.path.append(os.path.join(ROOT_DIR, 'mayavi'))
from kitti_util import *
def visualization():
import mayavi.mlab as mlab
dataset = kitti_object(r'./dataset/')
path = r'./dataset/testing/label_2/'
Save_Path = r'./save_3d_output/'
files = os.listdir(path)
for file in files:
name = file.split('.')[0]
save_path = Save_Path + name + '.png'
data_idx = int(name)
# Load data from dataset
objects = dataset.get_label_objects(data_idx)
img = dataset.get_image(data_idx)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
calib = dataset.get_calibration(data_idx)
print(' ------------ save image with 3D bounding box ------- ')
print('name:', name)
show_image_with_boxes(img, objects, calib, save_path, True)
if __name__=='__main__':
visualization()依賴代碼?kitti_util.py
# kitti_util.py
from __future__ import print_function
import os
import sys
import cv2
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt
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'))
class kitti_object(object):
def __init__(self, root_dir, split='testing'):
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)
self.image_dir = os.path.join(self.split_dir, 'image_2')
self.calib_dir = os.path.join(self.split_dir, 'calib')
self.label_dir = os.path.join(self.split_dir, 'label_2')
def __len__(self):
return self.num_samples
def get_image(self, idx):
assert(idx<self.num_samples)
img_filename = os.path.join(self.image_dir, '%06d.png'%(idx))
return load_image(img_filename)
def get_calibration(self, idx):
assert(idx<self.num_samples)
calib_filename = os.path.join(self.calib_dir, '%06d.txt'%(idx))
return Calibration(calib_filename)
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 read_label(label_filename)
def show_image_with_boxes(img, objects, calib, save_path, 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) # 畫2D框
box3d_pts_2d, box3d_pts_3d = compute_box_3d(obj, calib.P) # 獲取3D框-圖像(8*2)、3D框-相機坐標(biāo)系(8*3)
img2 = draw_projected_box3d(img2, box3d_pts_2d) # 在圖像上畫3D框
if show3d:
Image.fromarray(img2).save(save_path) # 保存帶有3D框的圖像
# Image.fromarray(img2).show()
else:
Image.fromarray(img1).save(save_path) # 保存帶有2D框的圖像
# Image.fromarray(img1).show()
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.'''
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)
def corners3d_to_img_boxes(self, corners3d):
"""
:param corners3d: (N, 8, 3) corners in rect coordinate
:return: boxes: (None, 4) [x1, y1, x2, y2] in rgb coordinate
:return: boxes_corner: (None, 8) [xi, yi] in rgb coordinate
"""
sample_num = corners3d.shape[0]
corners3d_hom = np.concatenate((corners3d, np.ones((sample_num, 8, 1))), axis=2) # (N, 8, 4)
img_pts = np.matmul(corners3d_hom, self.P.T) # (N, 8, 3)
x, y = img_pts[:, :, 0] / img_pts[:, :, 2], img_pts[:, :, 1] / img_pts[:, :, 2]
x1, y1 = np.min(x, axis=1), np.min(y, axis=1)
x2, y2 = np.max(x, axis=1), np.max(y, axis=1)
boxes = np.concatenate((x1.reshape(-1, 1), y1.reshape(-1, 1), x2.reshape(-1, 1), y2.reshape(-1, 1)), axis=1)
boxes_corner = np.concatenate((x.reshape(-1, 8, 1), y.reshape(-1, 8, 1)), axis=2)
return boxes, boxes_corner
# ===========================
# ------- 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):
'''
將3D坐標(biāo)點投影到圖像平面上，生成2D坐
pts_3d是一個nx3的矩陣，包含了待投影的3D坐標(biāo)點（每行一個點），P是相機的投影矩陣，通常是一個3x4的矩陣。
函數(shù)返回一個nx2的矩陣，包含了投影到圖像平面上的2D坐標(biāo)點。
'''
''' 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] # 獲取3D點的數(shù)量
pts_3d_extend = np.hstack((pts_3d, np.ones((n,1)))) # 將每個3D點的坐標(biāo)擴展為齊次坐標(biāo)形式（4D），通過在每個點的末尾添加1，創(chuàng)建了一個nx4的矩陣。
# print(('pts_3d_extend shape: ', pts_3d_extend.shape))
pts_2d = np.dot(pts_3d_extend, np.transpose(P)) # 將擴展的3D坐標(biāo)點矩陣與投影矩陣P相乘，得到一個nx3的矩陣，其中每一行包含了3D點在圖像平面上的投影坐標(biāo)。每個點的坐標(biāo)表示為[x, y, z]。
pts_2d[:,0] /= pts_2d[:,2] # 將投影坐標(biāo)中的x坐標(biāo)除以z坐標(biāo)，從而獲得2D圖像上的x坐標(biāo)。
pts_2d[:,1] /= pts_2d[:,2] # 將投影坐標(biāo)中的y坐標(biāo)除以z坐標(biāo)，從而獲得2D圖像上的y坐標(biāo)。
return pts_2d[:,0:2] # 返回一個nx2的矩陣,其中包含了每個3D點在2D圖像上的坐標(biāo)。
def compute_box_3d(obj, P):
'''
計算對象的3D邊界框在圖像平面上的投影
輸入: obj代表一個物體標(biāo)簽信息, P代表相機的投影矩陣-內(nèi)參。
輸出: 返回兩個值, corners_3d表示3D邊界框在 相機坐標(biāo)系 的8個角點的坐標(biāo)-3D坐標(biāo)。
corners_2d表示3D邊界框在 圖像上 的8個角點的坐標(biāo)-2D坐標(biāo)。
'''
# compute rotational matrix around yaw axis
# 計算一個繞Y軸旋轉(zhuǎn)的旋轉(zhuǎn)矩陣R，用于將3D坐標(biāo)從世界坐標(biāo)系轉(zhuǎn)換到相機坐標(biāo)系。obj.ry是對象的偏航角
R = roty(obj.ry)
# 3d bounding box dimensions
# 物體實際的長、寬、高
l = obj.l;
w = obj.w;
h = obj.h;
# 3d bounding box corners
# 存儲了3D邊界框的8個角點相對于對象中心的坐標(biāo)。這些坐標(biāo)定義了3D邊界框的形狀。
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
# 1、將3D邊界框的角點坐標(biāo)從對象坐標(biāo)系轉(zhuǎn)換到相機坐標(biāo)系。它使用了旋轉(zhuǎn)矩陣R
corners_3d = np.dot(R, np.vstack([x_corners,y_corners,z_corners]))
# 3D邊界框的坐標(biāo)進行平移
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];
# 2、檢查對象是否在相機前方，因為只有在相機前方的對象才會被繪制。
# 如果對象的Z坐標(biāo)（深度）小于0.1，就意味著對象在相機后方，那么corners_2d將被設(shè)置為None，函數(shù)將返回None。
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
# 3、將相機坐標(biāo)系下的3D邊界框的角點，投影到圖像平面上，得到它們在圖像上的2D坐標(biāo)。
corners_2d = project_to_image(np.transpose(corners_3d), P);
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=(0,60,255), thickness=2):
'''
qs: 包含8個3D邊界框角點坐標(biāo)的數(shù)組, 形狀為(8, 2)。圖像坐標(biāo)下的3D框, 8個頂點坐標(biāo)。
'''
''' Draw 3d bounding box in image
qs: (8,2) array of vertices for the 3d box in following order:
1 -------- 0
/| /|
2 -------- 3 .
| | | |
. 5 -------- 4
|/ |/
6 -------- 7
'''
qs = qs.astype(np.int32) # 將輸入的頂點坐標(biāo)轉(zhuǎn)換為整數(shù)類型，以便在圖像上繪制。
# 這個循環(huán)迭代4次，每次處理一個邊界框的一條邊。
for k in range(0,4):
# Ref: http://docs.enthought.com/mayavi/mayavi/auto/mlab_helper_functions.html
# 定義了要繪制的邊的起始點和結(jié)束點的索引。在這個循環(huán)中，它用于繪制邊界框的前四條邊。
i,j=k,(k+1)%4
cv2.line(image, (qs[i,0],qs[i,1]), (qs[j,0],qs[j,1]), color, thickness)
# 定義了要繪制的邊的起始點和結(jié)束點的索引。在這個循環(huán)中，它用于繪制邊界框的后四條邊，與前四條邊平行
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)
# 定義了要繪制的邊的起始點和結(jié)束點的索引。在這個循環(huán)中，它用于繪制連接前四條邊和后四條邊的邊界框的邊。
i,j=k,k+4
cv2.line(image, (qs[i,0],qs[i,1]), (qs[j,0],qs[j,1]), color, thickness)
return image
運行后會在save_3d_output中保存可視化的圖像。
模型推理結(jié)果可視化效果：
這個數(shù)據(jù)集的部分標(biāo)簽不準(zhǔn)確！??！
總結(jié)：有些失望，不準(zhǔn)確的標(biāo)簽占比較大；本來還想著用它替換Kitti的數(shù)據(jù)集。文章來源：http://www.zghlxwxcb.cn/news/detail-777478.html
只能用來做預(yù)訓(xùn)練，或者人工挑選標(biāo)簽做數(shù)據(jù)清洗。文章來源地址http://www.zghlxwxcb.cn/news/detail-777478.html
到了這里，關(guān)于3D檢測數(shù)據(jù)集 DAIR-V2X-V 轉(zhuǎn)為Kitti格式 | 可視化的文章就介紹完了。如果您還想了解更多內(nèi)容，請在右上角搜索TOY模板網(wǎng)以前的文章或繼續(xù)瀏覽下面的相關(guān)文章，希望大家以后多多支持TOY模板網(wǎng)！
