保定清苑住房和城乡建设局网站,云服务器如何搭建,网页打不开了,深圳全网营销网站2 Paddle3D 雷达点云CenterPoint模型训练–包含KITTI格式数据地址
2.0 数据集 百度DAIR-V2X开源路侧数据转kitti格式。
2.0.1 DAIR-V2X-I\velodyne中pcd格式的数据转为bin格式
参考源码#xff1a;雷达点云数据.pcd格式转.bin格式
def pcd2bin():import numpy as npimport…2 Paddle3D 雷达点云CenterPoint模型训练–包含KITTI格式数据地址
2.0 数据集 百度DAIR-V2X开源路侧数据转kitti格式。
2.0.1 DAIR-V2X-I\velodyne中pcd格式的数据转为bin格式
参考源码雷达点云数据.pcd格式转.bin格式
def pcd2bin():import numpy as npimport open3d as o3dfrom tqdm import tqdmimport ospcdPath rE:\DAIR-V2X-I\velodynebinPath rE:\DAIR-V2X-I\kitti\training\velodynefiles os.listdir(pcdPath)files [f for f in files if f[-4:].pcd]for ic in tqdm(range(len(files)), desc进度 ):f files[ic]pcdname os.path.join(pcdPath, f)binname os.path.join(binPath, f[:-4] .bin)# 读取PCD文件# pcd o3d.io.read_point_cloud(./data/002140.ply)pcd o3d.io.read_point_cloud(pcdname)# print(pcd\n, pcd)# print(pcd.points\n, pcd.points)points np.asarray(pcd.points)# print(points\n, points)# print(type(points))# print(points.shape)# 添加全0列point0 np.zeros((points.shape[0], 1))points np.column_stack((points,point0))# print(points.shape)# 查看点云图像# o3d.visualization.draw_geometries([pcd])# 将PCD格式保存为BIN格式使用.tofile实现# 理论上o3d.io.write_point_cloud也可以实现但是运行的时候没有报错但也并没有保存文件points.tofile(binname)o3d.io.write_point_cloud(os.path.join(binPath, f[:-4].bin), pcd) # if ic 1:# break可视化查看bin文件
def visBinData():可视化的形式查看点云数据的Bin文件:return:import numpy as npfrom tqdm import tqdmimport mayavi.mlabimport osbinPath rD:\lidar3D\data\Lidar0\bin# binPath r./datafiles os.listdir(binPath)files [f for f in files if f[-4:] .bin]for ic in tqdm(range(len(files)), desc进度 ):f files[ic]binname os.path.join(binPath, f)pointcloud np.fromfile(binname, dtypenp.float32, count-1).reshape([-1,4])x pointcloud[:, 0]y pointcloud[:, 1]z pointcloud[:, 2]r pointcloud[:, 3]d np.sqrt(x ** 2 y ** 2) # Map Distance from sensordegr np.degrees(np.arctan(z / d))vals heightif vals height:col zelse:col dfig mayavi.mlab.figure(bgcolor(0, 0, 0), size(640, 500))mayavi.mlab.points3d(x, y, z,col, # Values used for Colormodepoint,colormapspectral, # bone, copper, gnuplot# color(0, 1, 0), # Used a fixed (r,g,b) insteadfigurefig,)mayavi.mlab.show()breakdef visBinData():import open3d as o3dimport numpy as npimport os# 替换为你的 bin 文件路径bin_file_path rE:\DAIR-V2X-I\kitti_s\training\velodynefiles os.listdir(bin_file_path)for f in files[:]:# 读取 bin 文件bin_file os.path.join(bin_file_path, f)print(bin_file)points_np np.fromfile(bin_file, dtypenp.float32)print(points_np.shape)points_np points_np.reshape(-1, 4)print(points_np.shape)# 创建 Open3D 点云对象pcd o3d.geometry.PointCloud()pcd.points o3d.utility.Vector3dVector(points_np[:, :3])# 可视化点云o3d.visualization.draw_geometries([pcd])打印查看bin文件中的数据
def readbinfiles():import numpy as npprint(\n * * 10 myData * * 10)path rD:\lidar3D\data\mydatas1\kitti_my\training\velodyne\000003.bin# b1 np.fromfile(path, dtypenp.float)b2 np.fromfile(path, dtypenp.float32, count-1).reshape([-1, 4])print(type(b2))print(b2.shape)print(b2)2.0.2 先对应training中的所有数据准备好
————————training
______________calib
______________image_2
______________Label_2
______________velodyne【1】velodyne
这里是所有的.bin格式的点云文件
【2】iamge_2
这里是velodyne中的点云文件对应的图片文件。从原文件里面相同名称的图片复制过来就可以了这里原图片文件是.jpg的格式kitti里面查找.png的格式好像是写死的为避免麻烦可以先把图片直接重命名为.png的格式。
def jpg2png():import ospath rE:\DAIR-V2X-I\example71\training\image_2files os.listdir(path)for f in files:fsrc os.path.join(path, f)fdes os.path.join(path, f[:-4].png)os.rename(fsrc, fdes)【3】calib
两种方式一种方式是忽略DAIR-V2X数据中的标定直接复制kitti中calib/000000.txt if calibFlag:copy kitti-mini/training/calib/00000.txt cotent to my calibcalibPath rE:\DAIR-V2X-I\training\calibbinPath rE:\DAIR-V2X-I\training\velodynekitti_miniCalibPath rD:\lidar3D\data\kitti_mini\training\calib/000000.txtwith open(kitti_miniCalibPath, r) as calib_file:content calib_file.read()for f in os.listdir(binPath):with open(os.path.join(calibPath, f[:-4].txt), w) as wf:wf.write(content)另外一种方式是提取single-infrastructure-side\calib中的标定信息写进calib/xxxxxx.txt文件。因为DAIR-V2X只有一台相机写到p0: R0_rect 和 Tr_velo_to_cam 也从中读取。由于Tr_velo_to_cam 是固定的所以写死在了代码里面 if calibFlag1:V2XVelodynePath rE:\DAIR-V2X-I\training\velodynefiles os.listdir(V2XVelodynePath)files [f[:-4].txt for f in files]for f in files:content[]cameintriPath rE:\DAIR-V2X-I\single-infrastructure-side\calib\camera_intrinsicwith open(os.path.join(cameintriPath,f[:-4].json), r) as rf1:data json.load(rf1)p0 P0:for d in data[P]:p0 {}.format(d)R0_rect R0_rect:for r in data[R]:R0_rect {}.format(r)content.append(p0\n)content.append(P1: 1 0 0 0 0 1 0 0 0 0 1 0\n)content.append(P2: 1 0 0 0 0 1 0 0 0 0 1 0\n)content.append(P3: 1 0 0 0 0 1 0 0 0 0 1 0\n)content.append(R0_rect: 1 0 0 0 1 0 0 0 1\n)# vir2camePath os.path.join(V2XCalibPath, cliblist[1])# with open(os.path.join(vir2camePath,f[:-4].json), r) as rf1:# data json.load(rf1)# R data[rotation]# t data[translation]Tr_velo_to_cam Tr_velo_to_cam: -0.032055018882740139 -0.9974518923884874 0.020551248965447915 -2.190444561668236 -0.2240930139414797 0.002986041494130043 -0.8756800120708629 5.6360862566491909 0.9737455440255373 -0.041678350017788 -0.2023375046787095 1.4163664770754852\ncontent.append(Tr_velo_to_cam)content.append(Tr_imu_to_velo: 1 0 0 0 0 1 0 0 0 0 1 0)fcalibname os.path.join(mykittiCalibPath, f)# content[-1] content[-1].replace(\n, )with open(fcalibname, w, encodingutf-8) as wf:wf.writelines(content)【4】Label_2
方法{{{见 《 2.1.3中——【4】问题记录3 —— 重新生成label_2文件过程如下》章节}}}
点云的标签文件single-infrastructure-side\label\virtuallidar*.json转为txt文件存储这里3Class故只取类型[“Car”, “Cyclist”, “Pedestrian”][‘Cyclist’, ‘Car’, ‘Truck’]。
这里将每个标签目标记作temp temp [0:16] 共计16列含义分别为
类别[0]是否截断[1]是否遮挡[2]观察角度[3]图像左上右下[4:8]高宽长[8:11]相机坐标系xyz[11:14]方向角[14]置信度[15]
注意测试数据集才有最后一列的置信度。
json转txt代码如下因为只训练点云数据所以与图片相关的值这里用0代替。
如下代码是直接取json文件中的数据尺寸取雷达坐标系下的位置取相机坐标系下的。 if label_2Flag:# 1 创建文件夹 label_2label_2Path rE:\DAIR-V2X-I\kitti_s\training\label_22_kitti3clsNewXYZjudgefolderExit_mkdir(label_2Path)V2XVelodynePath rE:\DAIR-V2X-I\kitti_s\training\velodynefiles os.listdir(V2XVelodynePath)files [f[:-4] .txt for f in files]V2XLabelPathcamera rE:\DAIR-V2X-I\single-infrastructure-side\label\cameraV2XLabelPathlidar rE:\DAIR-V2X-I\single-infrastructure-side\label\virtuallidar# 2 转化生成对应的.txt标签文件labelsList {}labelsListnew {}for f in files[:]:# print(f)with open(os.path.join(V2XLabelPathcamera, f[:-4] .json), r) as rf1:dataca json.load(rf1)with open(os.path.join(V2XLabelPathlidar, f[:-4] .json), r) as rf1:datali json.load(rf1)label_list []if len(dataca) ! len(datali):print(Error: len(dataca) ! len(datali))continuefor oi in range(len(dataca)):obj_ca dataca[oi]obj_li datali[oi]label_name obj_ca[type]label_nameli obj_li[type]if label_name ! label_nameli:print(Error: label_name ! label_nameli)continue# static labelsif label_name in labelsList.keys():labelsList[label_name] 1else:labelsList[label_name] 1# updata label type# Car 、 Cyclist、if label_name Trafficcone or label_name ScooterRider or label_name Barrowlist:continueelif label_name Motorcyclist:label_name Cyclistelif label_name Van:label_name Carelif label_name Bus:label_name Carelif label_name Truck:label_name Carif label_name in labelsListnew.keys():labelsListnew[label_name] 1else:labelsListnew[label_name] 1scale obj_li[3d_dimensions] # 尺寸取雷达坐标系下的pos obj_ca[3d_location] # 位置取相机坐标系下的rot obj_li[rotation]alpha obj_ca[alpha] # 观察角度[3]cabox obj_ca[2d_box] # 图像左上右下[4:8]occluded_state obj_ca[occluded_state] # 是否遮挡[2]truncated_state obj_ca[truncated_state] # 是否截断[1]tempFlag True # label_list.append(temp)if tempFlag True:# 2.1 temp 追加目标类型# temp [0:16] 共计16列# 类别[0]是否截断[1]是否遮挡[2]观察角度[3]图像左上右下[4:8]高宽长[8:11]# 相机坐标系xyz[11:14]方向角[14]置信度[15]# 注意测试数据集才有最后一列的置信度temp label_name truncated_state occluded_state 0 \cabox[xmin] cabox[ymin] cabox[xmax] cabox[ymax] temp (scale[h].split(.)[0] . scale[h].split(.)[1][:2] )temp (scale[w].split(.)[0] . scale[w].split(.)[1][:2] )temp (scale[l].split(.)[0] . scale[l].split(.)[1][:2] )# 2.2.1 pos_xyz# temp (pos[x].split(.)[0] . pos[x].split(.)[1][:2] )# temp (pos[y].split(.)[0] . pos[y].split(.)[1][:2] )# temp (pos[z].split(.)[0] . pos[z].split(.)[1][:2] )# 2.2 固定使用转换矩阵lidar_to_cam后# 此处的矩阵目的将lidar转到cam下再转换到rect下同样追加到temp中(temp追加pos_xyz)# trans_Mat: calib中的 Tr_velo_to_camtrans_Mat np.array([[-3.205501888274e-02, -9.974518923885e-01, 2.055124896545e-02, -2.190444561668e00],[-2.240930139415e-01, 2.986041494130e-03, -8.756800120709e-01, 5.636086256649e00],[9.737455440255e-01, -4.167835001779e-02, -2.023375046787e-01, 1.416366477075e00],[0, 0, 0, 1]])# R0_rect: 三维单位矩阵rect_Mat np.array([[1, 0, 0, 0],[0, 1, 0, 0],[0, 0, 1, 0],[0, 0, 0, 1]])ptx str2float(pos[x])pty str2float(pos[y])# print(pos[z], scale[l], type(scale[l])\n, pos[z], scale[l], type(scale[l]))ptz str2float(pos[z]) - (0.5 * str2float(scale[l]))# lidar坐标系下的相机中心点pt_in_lidar np.array([[ptx],[pty],[ptz],[1.]])pt_in_camera np.matmul(trans_Mat, pt_in_lidar)pt_in_rect np.matmul(rect_Mat, pt_in_camera)temp str(pt_in_rect[0, 0]).split(.)[0] . str(pt_in_rect[0, 0]).split(.)[1][0:2]temp temp str(pt_in_rect[1, 0]).split(.)[0] . str(pt_in_rect[1, 0]).split(.)[1][0:2]temp temp str(pt_in_rect[2, 0]).split(.)[0] . str(pt_in_rect[2, 0]).split(.)[1][0:2]temp ## 2.3 temp 追加 rot_xyz(先将 rot 航向角 返回到0-360°之间)rot - str2float(rot) - (np.pi / 2)if rot np.pi:rot rot - 2 * np.pielif rot -np.pi:rot rot 2 * np.pitemp str(rot).split(.)[0] . str(rot).split(.)[1][0:2] \nlabel_list.append(temp)label_list[-1] label_list[-1].replace(\n, )# print(label_list)with open(os.path.join(label_2Path, f[:-4].txt), w) as wf:wf.writelines(label_list)print(labels Statics , labelsList)print(labels Statics , labelsListnew)如下代码考虑重新计算了位置信息固定使用转换矩阵lidar_to_cam。[“Car”, “Cyclist”, “Pedestrian”] if label_2Flag:# 1 创建文件夹 label_2label_2Path rE:\DAIR-V2X-I\kitti\training\Label_2judgefolderExit_mkdir(label_2Path)V2XVelodynePath rE:\DAIR-V2X-I\kitti\training\velodynefiles os.listdir(V2XVelodynePath)files [f[:-4] .txt for f in files]V2XLabelPathcamera rE:\DAIR-V2X-I\single-infrastructure-side\label\cameraV2XLabelPathlidar rE:\DAIR-V2X-I\single-infrastructure-side\label\virtuallidar# 2 转化生成对应的.txt标签文件labelsList {}labelsListnew {}for f in files[:]:# print(f)with open(os.path.join(V2XLabelPathcamera, f[:-4] .json), r) as rf1:dataca json.load(rf1)with open(os.path.join(V2XLabelPathlidar, f[:-4] .json), r) as rf1:datali json.load(rf1)label_list []if len(dataca) ! len(datali):print(Error: len(dataca) ! len(datali))continuefor oi in range(len(dataca)):obj_ca dataca[oi]obj_li datali[oi]label_name obj_ca[type]label_nameli obj_li[type]if label_name ! label_nameli:print(Error: label_name ! label_nameli)continue# static labelsif label_name in labelsList.keys():labelsList[label_name] 1else:labelsList[label_name] 1# updata label type# Car 、 Cyclist、if label_name Trafficcone or label_name ScooterRider or label_name Barrowlist:continueelif label_name Motorcyclist:label_name Cyclistelif label_name Van:label_name Carelif label_name Bus:label_name Carelif label_name Truck:label_name Carif label_name in labelsListnew.keys():labelsListnew[label_name] 1else:labelsListnew[label_name] 1scale obj_li[3d_dimensions] # 尺寸取雷达坐标系下的pos obj_ca[3d_location] # 位置取相机坐标系下的rot obj_li[rotation]tempFlag False # label_list.append(temp)if tempFlag True:# 2.1 temp 追加目标类型# temp [0:16] 共计16列# 类别[0]是否截断[1]是否遮挡[2]观察角度[3]图像左上右下[4:8]高宽长[8:11]# 相机坐标系xyz[11:14]方向角[14]置信度[15]# 注意测试数据集才有最后一列的置信度temp label_name 0 0 0 0 0 0 0 temp (scale[h].split(.)[0] . scale[h].split(.)[1][:2] )temp (scale[w].split(.)[0] . scale[w].split(.)[1][:2] )temp (scale[l].split(.)[0] . scale[l].split(.)[1][:2] )# 2.2.1 pos_xyz# temp (pos[x].split(.)[0] . pos[x].split(.)[1][:2] )# temp (pos[y].split(.)[0] . pos[y].split(.)[1][:2] )# temp (pos[z].split(.)[0] . pos[z].split(.)[1][:2] )# 2.2 固定使用转换矩阵lidar_to_cam后# 此处的矩阵目的将lidar转到cam下再转换到rect下同样追加到temp中(temp追加pos_xyz)trans_Mat np.array([[6.927964000000e-03, -9.999722000000e-01, -2.757829000000e-03, -2.457729000000e-02],[-1.162982000000e-03, 2.749836000000e-03, -9.999955000000e-01, -6.127237000000e-02],[9.999753000000e-01, 6.931141000000e-03, -1.143899000000e-03, -3.321029000000e-01],[0, 0, 0, 1]])rect_Mat np.array([[9.999128000000e-01, 1.009263000000e-02, -8.511932000000e-03, 0],[-1.012729000000e-02, 9.999406000000e-01, -4.037671000000e-03, 0],[8.470675000000e-03, 4.123522000000e-03, 9.999556000000e-01, 0],[0, 0, 0, 1]])ptx str2float(pos[x])pty str2float(pos[y])# print(pos[z], scale[l], type(scale[l])\n, pos[z], scale[l], type(scale[l]))ptz str2float(pos[z]) - (0.5 * str2float(scale[l]))# lidar坐标系下的相机中心点pt_in_lidar np.array([[ptx],[pty],[ptz],[1.]])pt_in_camera np.matmul(trans_Mat, pt_in_lidar)pt_in_rect np.matmul(rect_Mat, pt_in_camera)temp str(pt_in_rect[0, 0]).split(.)[0] . str(pt_in_rect[0, 0]).split(.)[1][0:2]temp temp str(pt_in_rect[1, 0]).split(.)[0] . str(pt_in_rect[1, 0]).split(.)[1][0:2]temp temp str(pt_in_rect[2, 0]).split(.)[0] . str(pt_in_rect[2, 0]).split(.)[1][0:2]temp ## 2.3 temp 追加 rot_xyz(先将 rot 航向角 返回到0-360°之间)rot - str2float(rot) - (np.pi / 2)if rot np.pi:rot rot - 2 * np.pielif rot -np.pi:rot rot 2 * np.pitemp str(rot).split(.)[0] . str(rot).split(.)[1][0:2] \nlabel_list.append(temp)# label_list[-1] label_list[-1].replace(\n, )# with open(os.path.join(label_2Path, f[:-4].txt), w) as wf:# wf.writelines(label_list)print(labels Statics , labelsList )print(labels Statics , labelsListnew )2.0.3 kitti\testing中的数据类似training中数据
仅需移动部分数据作为test是数据即可
————————testing
______________calib
______________image_2
______________Label_2
______________velodyne2.0.4 写kitti/ImageSets 中的txt文件
注意如下代码中val.txt 与test.txt中的文件是一致的
def filesPath2txt():import os, randompathtrain E:\DAIR-V2X-I\kitti_s/training/velodyne/pathtest E:\DAIR-V2X-I\kitti_s/testing/velodyne/txtPath E:\DAIR-V2X-I\kitti_s/ImageSets/txt_path_train os.path.join(txtPath, train.txt)txt_path_val os.path.join(txtPath, val.txt)# txt_path_trainval os.path.join(txtPath, trainval.txt)txt_path_test os.path.join(txtPath, test.txt)# mkdirImageSets(txt_path_train, txt_path_val, txt_path_trainval, txt_path_test)files os.listdir(pathtrain)filesList [f[:-4] \n for f in files]filestr filesListfilestr[-1] filestr[-1].replace(\n, )with open(txt_path_train, w, encodingutf-8) as wf:wf.writelines(filestr)files os.listdir(pathtest)filesList [f[:-4] \n for f in files]filesval filesList # random.sample(filesList, 700)filesval[-1] filesval[-1].replace(\n, )with open(txt_path_val, w, encodingutf-8) as wf:wf.writelines(filesval)# filestest random.sample(filesList, 400)# filestest[-1] filestest[-1].replace(\n, )with open(txt_path_test, w, encodingutf-8) as wf:wf.writelines(filesval)2.1 paddle3D训练
cd ./Paddle3D2.1.1 数据
【1】数据
数据存放在Paddle3D/datasets目录下结构如下
datasets/
datasets/KITTI/
————datasets/KITTI/ImageSets
————datasets/KITTI/testing
————datasets/KITTI/training【2】数据预处理
使用如下代码完成数据的预处理操作
python tools/create_det_gt_database.py --dataset_name kitti --dataset_root ./datasets/KITTI --save_dir ./datasets/KITTI上述过程打印如下运行结束会生成datasets/KITTI/kitti_train_gt_database目录。
root/anaconda3/envs/pip_paddle_env/lib/python3.8/site-packages/setuptools/sandbox.py:13: DeprecationWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.htmlimport pkg_resources
root/anaconda3/envs/pip_paddle_env/lib/python3.8/site-packages/pkg_resources/__init__.py:2871: DeprecationWarning: Deprecated call to pkg_resources.declare_namespace(mpl_toolkits).
Implementing implicit namespace packages (as specified in PEP 420) is preferred to pkg_resources.declare_namespace. See https://setuptools.pypa.io/en/latest/references/keywords.html#keyword-namespace-packagesdeclare_namespace(pkg)
root/anaconda3/envs/pip_paddle_env/lib/python3.8/site-packages/pkg_resources/__init__.py:2871: DeprecationWarning: Deprecated call to pkg_resources.declare_namespace(google).
Implementing implicit namespace packages (as specified in PEP 420) is preferred to pkg_resources.declare_namespace. See https://setuptools.pypa.io/en/latest/references/keywords.html#keyword-namespace-packagesdeclare_namespace(pkg)
ortools not installed, install it by pip install ortools9.1.9490 if you run BEVLaneDet model
2023-12-26 17:45:46,823 - INFO - Begin to generate a database for the KITTI dataset.
2023-12-26 17:46:06,774 - INFO - [##################################################] 100.00%
2023-12-26 17:46:07,012 - INFO - The database generation has been done.2.1.2 模型配置文件
为避免修改原模型配置文件先复制一份并命名为centerpoint_pillars_016voxel_kitti_my.yml
cp ./configs/centerpoint/centerpoint_pillars_016voxel_kitti.yml ./configs/centerpoint/centerpoint_pillars_016voxel_kitti_my.yml核对文件中的相关配置信息
train_dataset:type: KittiPCDatasetdataset_root: datasets/KITTI... ...class_names: [Car, Cyclist, Pedestrian]2.1.3 训练流程及问题调试
【1】使用如下代码进行训练
# python -m paddle.distributed.launch --gpus 0,1,2,3,4,5,6,7 tools/train.py --config configs/centerpoint/centerpoint_pillars_016voxel_kitti.yml --save_dir ./output_kitti --num_workers 4 --save_interval 5python tools/train.py --config configs/centerpoint/centerpoint_pillars_016voxel_kitti_my.yml --save_dir ./output_kitti --save_interval 5 112.log参数介绍
-m使用python -m paddle.distributed.launch方法启动分布式训练任务。
参考https://www.paddlepaddle.org.cn/documentation/docs/zh/api/paddle/distributed/launch_cn.html112.log : 将其中的打印保存在112.log中【2】问题记录1
【2-1】SystemError: (Fatal) Blocking queue is killed because the data reader raises an exception.
【2-2】KeyError: ‘DataLoader worker(1) caught KeyError with message:\nTraceback (most recent call last):\n File “/home/… …”… … self.sampler_per_class[cls_name].sampling(num_samples)\n KeyError: ‘Car’\n’
【2-1、2-2】解决方法
参考1SystemError: (Fatal) Blocking queue is killed because the data reader raises an exception 没找到源码中的相关位置。
参考2SystemError: (Fatal) Blocking queue is killed (baidu.com)
利用“参考2”中提供的方法, 训练的时候将num_worker设置为0。如下
python tools/train.py --config configs/centerpoint/centerpoint_pillars_016voxel_kitti_my.yml --save_dir ./output_kitti --num_workers 0 --save_interval 5将num_workers 设置为0训练的时候报如下错误
/Paddle3D/paddle3d/transforms/sampling.py, line 172, in samplingsampling_annos self.sampler_per_class[cls_name].sampling(num_samples)
KeyError: Car采用如下代码打印/Paddle3D/paddle3d/transforms/sampling.py, line 172中的信息结果是{}。
print(self.sampler_per_class, self.sampler_per_class)
》》》》》》self.sampler_per_class {}发现是配置文件中class_balanced_sampling参数是设置为False的原因。修改class_balanced_sampling为True如下
train_dataset:type: KittiPCDatasetdataset_root: datasets/KITTI... ...mode: trainclass_balanced_sampling: Trueclass_names: [Car, Cyclist, Pedestrian]【3】问题记录2
【3-1】ZeroDivisionError: float division by zero
详细报错包含如下
File /home/mec/hulijuan/Paddle3D/paddle3d/datasets/kitti/kitti_det.py, line 89, in listcompsampling_ratios [balanced_frac / frac for frac in fracs]
ZeroDivisionError: float division by zero上述问题打印 print(‘kitti_det.py’, cls_dist) 发现是自己数据集中没有Pedestrian类别而class_names中包含该类别。解决办法是增加包含 Pedestrian类 的数据。
【4】问题记录3
【4-1】又报错问题同“【2】问题记录1 SystemError: … KeyError: …”只能是大概上面的问题没有从根本上解决掉
首先看KeyError的问题。
在paddle3d/transforms/sampling.py/class SamplingDatabase(TransformABC):/def __ init __()中增加打印。如下 def __init__(self,min_num_points_in_box_per_class: Dict[str, int],max_num_samples_per_class: Dict[str, int],database_anno_path: str,database_root: str,class_names: List[str],ignored_difficulty: List[int] None):self.min_num_points_in_box_per_class min_num_points_in_box_per_classself.max_num_samples_per_class max_num_samples_per_classself.database_anno_path database_anno_pathwith open(database_anno_path, rb) as f:database_anno pickle.load(f)print(sampling.py__line58~~~~~~~~~~~~~~~~~~~~58database_anno: , database_anno)if not osp.exists(database_root):raise ValueError(fDatabase root path {database_root} does not exist!!!)self.database_root database_rootself.class_names class_namesself.database_anno self._filter_min_num_points_in_box(database_anno)self.ignored_difficulty ignored_difficultyif ignored_difficulty is not None:self.database_anno self._filter_ignored_difficulty(self.database_anno)self.sampler_per_class dict()print(sampling.py__line70~~~~~~~~~~~~~~~70database_anno: , self.database_anno)for cls_name, annos in self.database_anno.items():self.sampler_per_class[cls_name] Sampler(cls_name, annos)通过打印可以看出sampling.py–line70的打印是空字典而sampling.py–line58的打印部分如下
上图可以看出num_points_in_box 的值为 0。导致如下代码运行后database_anno 变成了空字典
self.database_anno self._filter_min_num_points_in_box(database_anno)上述问题应该是针对点云文件生成标签文件的时候方法错了。
重新生成label_2文件过程如下
1 创建环境
conda create -n pcd2bin_env python3.8
2 激活环境
conda activate pcd2bin_env
3 安装pypcd
3.1 参考https://blog.csdn.net/weixin_44450684/article/details/92812746
如下流程
git clone https://github.com/dimatura/pypcd
cd pypcd
git fetch origin pull/9/head:python3
git checkout python3
python3 setup.py install --user
python3
from pypcd import pypcd
pc pypcd.PointCloud.from_path(pointcloud.pcd)
利用开源程序重新生成label_2文件过程如下
源码百度网盘地址执行程序
2.2 数据集地址KITTI格式DAIR-V2X-I7058帧数据长期有效
大小22G
链接https://pan.baidu.com/s/1gG_S6Vtx4iWAfAVAfOxWWw
提取码p48l
数据没有问题 但是label_2中的XYZ需要根据前述lanel_2的方法重新生成。需下载该数据集的标签文件single-infrastructure-side-json.zip2.3 模型
2.3.1 模型评估
2.3.2 模型测试
2.3.3 导出模型
