OpenPCDet 训练自己的数据集详细教程！

---

前言

这些天一直在尝试通过OpenPCDet平台训练自己的数据集（非kitti格式），好在最后终于跑通了，特此记录一下训练过程。

---

一、pcd转bin

笔者自己的点云数据是pcd格式的，参照kitti训练过程是需要转成bin格式的。
下面给出转换代码：

# -*- coding: utf-8 -*- 
# @Time : 2022/7/25 11:30 
# @Author : JulyLi
# @File : pcd2bin.py

import numpy as np
import os
import argparse
from pypcd import pypcd
import csv
from tqdm import tqdm


def main():
    ## Add parser
    parser = argparse.ArgumentParser(description="Convert .pcd to .bin")
    parser.add_argument(
        "--pcd_path",
        help=".pcd file path.",
        type=str,
        default="pcd_raw1"
    )
    parser.add_argument(
        "--bin_path",
        help=".bin file path.",
        type=str,
        default="bin"
    )
    parser.add_argument(
        "--file_name",
        help="File name.",
        type=str,
        default="file_name"
    )
    args = parser.parse_args()

    ## Find all pcd files
    pcd_files = []
    for (path, dir, files) in os.walk(args.pcd_path):
        for filename in files:
            # print(filename)
            ext = os.path.splitext(filename)[-1]
            if ext == '.pcd':
                pcd_files.append(path + "/" + filename)

    ## Sort pcd files by file name
    pcd_files.sort()
    print("Finish to load point clouds!")

    ## Make bin_path directory
    try:
        if not (os.path.isdir(args.bin_path)):
            os.makedirs(os.path.join(args.bin_path))
    except OSError as e:
        # if e.errno != errno.EEXIST:
        #     print("Failed to create directory!!!!!")
            raise

    ## Generate csv meta file
    csv_file_path = os.path.join(args.bin_path, "meta.csv")
    csv_file = open(csv_file_path, "w")
    meta_file = csv.writer(
        csv_file, delimiter=",", quotechar="|", quoting=csv.QUOTE_MINIMAL
    )
    ## Write csv meta file header
    meta_file.writerow(
        [
            "pcd file name",
            "bin file name",
        ]
    )
    print("Finish to generate csv meta file")

    ## Converting Process
    print("Converting Start!")
    seq = 0
    for pcd_file in tqdm(pcd_files):
        ## Get pcd file
        pc = pypcd.PointCloud.from_path(pcd_file)

        ## Generate bin file name
        # bin_file_name = "{}_{:05d}.bin".format(args.file_name, seq)
        bin_file_name = "{:05d}.bin".format(seq)
        bin_file_path = os.path.join(args.bin_path, bin_file_name)

        ## Get data from pcd (x, y, z, intensity, ring, time)
        np_x = (np.array(pc.pc_data['x'], dtype=np.float32)).astype(np.float32)
        np_y = (np.array(pc.pc_data['y'], dtype=np.float32)).astype(np.float32)
        np_z = (np.array(pc.pc_data['z'], dtype=np.float32)).astype(np.float32)
        np_i = (np.array(pc.pc_data['intensity'], dtype=np.float32)).astype(np.float32) / 256
        # np_r = (np.array(pc.pc_data['ring'], dtype=np.float32)).astype(np.float32)
        # np_t = (np.array(pc.pc_data['time'], dtype=np.float32)).astype(np.float32)

        ## Stack all data
        points_32 = np.transpose(np.vstack((np_x, np_y, np_z, np_i)))

        ## Save bin file
        points_32.tofile(bin_file_path)

        ## Write csv meta file
        meta_file.writerow(
            [os.path.split(pcd_file)[-1], bin_file_name]
        )

        seq = seq + 1


if __name__ == "__main__":
    main()

二、labelCloud 工具安装与使用

拉取源码

git clone https://github.com/ch-sa/labelCloud.git

安装依赖

pip install -r requirements.txt

启动程序

python labelCloud.py

启动后出现如下界面：

在setting界面按需设置，笔者这里按kitti格式生成label数据：

标注完成后会在对应目录下生成标签：

标签内容大致如下：

三、训练

仿写代码

把pcdet/datasets/kitti文件夹复制并改名为pcdet/datasets/custom，然后把pcdet/utils/object3d_kitti.py复制为pcdet/utils/object3d_custom.py
把data/kitti文件夹复制并改名为data/custom，然后修改训练信息，结构如下：

custom
├── ImageSets
│   ├── test.txt
│   ├── train.txt
├── testing
│   ├── velodyne
├── training
│   ├── label_2
│   ├── velodyne

对pcdet/datasets/custom/custom_dataset.py进行改写

import copy
import pickle
import os

import numpy as np
from skimage import io

from . import custom_utils
from ...ops.roiaware_pool3d import roiaware_pool3d_utils
from ...utils import box_utils, common_utils, object3d_custom
from ..dataset import DatasetTemplate

class CustomDataset(DatasetTemplate):
    def __init__(self, dataset_cfg, class_names, training=True, root_path=None, logger=None, ext='.bin'):
        """
        Args:
            root_path:
            dataset_cfg:
            class_names:
            training:
            logger:
        """
        super().__init__(
            dataset_cfg=dataset_cfg, class_names=class_names, training=training, root_path=root_path, logger=logger
        )
        self.split = self.dataset_cfg.DATA_SPLIT[self.mode]
        self.root_split_path = os.path.join(self.root_path, ('training' if self.split != 'test' else 'testing'))

        split_dir = os.path.join(self.root_path, 'ImageSets',(self.split + '.txt'))
        self.sample_id_list = [x.strip() for x in open(split_dir).readlines()] if os.path.exists(split_dir) else None

        self.custom_infos = []
        self.include_custom_data(self.mode)
        self.ext = ext


    def include_custom_data(self, mode):
        if self.logger is not None:
            self.logger.info('Loading Custom dataset.')
        custom_infos = []

        for info_path in self.dataset_cfg.INFO_PATH[mode]:
            info_path = self.root_path / info_path
            if not info_path.exists():
                continue
            with open(info_path, 'rb') as f:
                infos = pickle.load(f)
                custom_infos.extend(infos)
        
        self.custom_infos.extend(custom_infos)

        if self.logger is not None:
            self.logger.info('Total samples for CUSTOM dataset: %d' % (len(custom_infos)))
    

    def get_infos(self, num_workers=16, has_label=True, count_inside_pts=True, sample_id_list=None):
        import concurrent.futures as futures

        # Process single scene
        def process_single_scene(sample_idx):
            print('%s sample_idx: %s' % (self.split, sample_idx))
            # define an empty dict
            info = {}
            # pts infos: dimention and idx
            pc_info = {'num_features': 4, 'lidar_idx': sample_idx}
            # add to pts infos
            info['point_cloud'] = pc_info

            # no images, calibs are need to transform the labels

            type_to_id = {'Car': 1, 'Pedestrian': 2, 'Cyclist': 3}
            if has_label:
                # read labels to build object list according to idx
                obj_list = self.get_label(sample_idx)
                # build an empty annotations dict
                annotations = {}
                # add to annotations ==> refer to 'object3d_custom' (no truncated,occluded,alpha,bbox)
                annotations['name'] = np.array([obj.cls_type for obj in obj_list]) # 1-dimension
                # hwl(camera) format 2-dimension: The kitti-labels are in camera-coord
                # h,w,l -> 0.21,0.22,0.33 (see object3d_custom.py h=label[8], w=label[9], l=label[10])
                annotations['dimensions'] = np.array([[obj.l, obj.h, obj.w] for obj in obj_list])             
                annotations['location'] = np.concatenate([obj.loc.reshape(1,3) for obj in obj_list])
                annotations['rotation_y'] = np.array([obj.ry for obj in obj_list]) # 1-dimension

                num_objects = len([obj.cls_type for obj in obj_list if obj.cls_type != 'DontCare'])
                num_gt = len(annotations['name'])
                index = list(range(num_objects)) + [-1] * (num_gt - num_objects)
                annotations['index'] = np.array(index, dtype=np.int32)

                loc = annotations['location'][:num_objects]
                dims = annotations['dimensions'][:num_objects]
                rots = annotations['rotation_y'][:num_objects]
                # camera -> lidar: The points of custom_dataset are already in lidar-coord
                # But the labels are in camera-coord and need to transform
                loc_lidar = self.get_calib(loc)
                l, h, w = dims[:, 0:1], dims[:, 1:2], dims[:, 2:3]
                # bottom center -> object center: no need for loc_lidar[:, 2] += h[:, 0] / 2
                # print("sample_idx: ", sample_idx, "loc: ", loc, "loc_lidar: " , sample_idx, loc_lidar)
                # get gt_boxes_lidar see https://zhuanlan.zhihu.com/p/152120636
                gt_boxes_lidar = np.concatenate([loc_lidar, l, w, h, (np.pi / 2 - rots[..., np.newaxis])], axis=1) # 2-dimension array
                annotations['gt_boxes_lidar'] = gt_boxes_lidar
                
                # add annotation info
                info['annos'] = annotations
            
            return info
        
        sample_id_list = sample_id_list if sample_id_list is not None else self.sample_id_list
        # create a thread pool to improve the velocity
        with futures.ThreadPoolExecutor(num_workers) as executor:
            infos = executor.map(process_single_scene, sample_id_list)
        # infos is a list that each element represents per frame
        return list(infos)
                

    def get_calib(self, loc):
        """
        This calibration is different from the kitti dataset.
        The transform formual of labelCloud: ROOT/labelCloud/io/labels/kitti.py: import labels
            if self.transformed:
                centroid = centroid[2], -centroid[0], centroid[1] - 2.3
            dimensions = [float(v) for v in line_elements[8:11]]
            if self.transformed:
                dimensions = dimensions[2], dimensions[1], dimensions[0]
            bbox = BBox(*centroid, *dimensions)
        """
        loc_lidar = np.concatenate([np.array((float(loc_obj[2]), float(-loc_obj[0]), float(loc_obj[1]-2.3)), dtype=np.float32).reshape(1,3) for loc_obj in loc])
        return loc_lidar
                

    def get_label(self, idx):
        # get labels
        label_file = self.root_split_path / 'label_2' / ('%s.txt' % idx)
        assert label_file.exists()
        return object3d_custom.get_objects_from_label(label_file)


    def get_lidar(self, idx, getitem):
        """
            Loads point clouds for a sample
                Args:
                    index (int): Index of the point cloud file to get.
                Returns:
                    np.array(N, 4): point cloud.
        """
        # get lidar statistics
        if getitem == True:
            lidar_file = self.root_split_path + '/velodyne/' + ('%s.bin' % idx)
        else:
            lidar_file = self.root_split_path / 'velodyne' / ('%s.bin' % idx)
        return np.fromfile(str(lidar_file), dtype=np.float32).reshape(-1, 4)


    def set_split(self, split):
        super().__init__(
            dataset_cfg=self.dataset_cfg, class_names=self.class_names, training=self.training, root_path=self.root_path, logger=self.logger
        )
        self.split = split
        self.root_split_path = self.root_path / ('training' if self.split != 'test' else 'testing')

        split_dir = self.root_path / 'ImageSets' / (self.split + '.txt')
        self.sample_id_list = [x.strip() for x in open(split_dir).readlines()] if split_dir.exists() else None


    # Create gt database for data augmentation
    def create_groundtruth_database(self, info_path=None, used_classes=None, split='train'):
        import torch

        # Specify the direction
        database_save_path = Path(self.root_path) / ('gt_database' if split == 'train' else ('gt_database_%s' % split))
        db_info_save_path = Path(self.root_path) / ('custom_dbinfos_%s.pkl' % split)

        database_save_path.mkdir(parents=True, exist_ok=True)
        all_db_infos = {}

        # Open 'custom_train_info.pkl'
        with open(info_path, 'rb') as f:
            infos = pickle.load(f)

        # For each .bin file
        for k in range(len(infos)):
            print('gt_database sample: %d/%d' % (k + 1, len(infos)))
            # Get current scene info
            info = infos[k]
            sample_idx = info['point_cloud']['lidar_idx']
            points = self.get_lidar(sample_idx, False)
            annos = info['annos']
            names = annos['name']
            gt_boxes = annos['gt_boxes_lidar']

            num_obj = gt_boxes.shape[0]
            point_indices = roiaware_pool3d_utils.points_in_boxes_cpu(
                torch.from_numpy(points[:, 0:3]), torch.from_numpy(gt_boxes)
            ).numpy()  # (nboxes, npoints)

            for i in range(num_obj):
                filename = '%s_%s_%d.bin' % (sample_idx, names[i], i)
                filepath = database_save_path / filename
                gt_points = points[point_indices[i] > 0]

                gt_points[:, :3] -= gt_boxes[i, :3]
                with open(filepath, 'w') as f:
                    gt_points.tofile(f)

                if (used_classes is None) or names[i] in used_classes:
                    db_path = str(filepath.relative_to(self.root_path))  # gt_database/xxxxx.bin
                    db_info = {'name': names[i], 'path': db_path, 'gt_idx': i,
                               'box3d_lidar': gt_boxes[i], 'num_points_in_gt': gt_points.shape[0]}
                    if names[i] in all_db_infos:
                        all_db_infos[names[i]].append(db_info)
                    else:
                        all_db_infos[names[i]] = [db_info]

        # Output the num of all classes in database
        for k, v in all_db_infos.items():
            print('Database %s: %d' % (k, len(v)))

        with open(db_info_save_path, 'wb') as f:
            pickle.dump(all_db_infos, f)

    @staticmethod
    def generate_prediction_dicts(batch_dict, pred_dicts, class_names, output_path=None):
        """
        Args:
            batch_dict:
                frame_id:
            pred_dicts: list of pred_dicts
                pred_boxes: (N,7), Tensor
                pred_scores: (N), Tensor
                pred_lables: (N), Tensor
            class_names:
            output_path:
        Returns:
        """
        def get_template_prediction(num_smaples):
            ret_dict = {
                'name': np.zeros(num_smaples), 'alpha' : np.zeros(num_smaples),
                'dimensions': np.zeros([num_smaples, 3]), 'location': np.zeros([num_smaples, 3]),
                'rotation_y': np.zero(num_smaples), 'score': np.zeros(num_smaples),
                'boxes_lidar': np.zeros([num_smaples, 7])
            }
            return ret_dict

        def generate_single_sample_dict(batch_index, box_dict):
            pred_scores = box_dict['pred_scores'].cpu().numpy()
            pred_boxes = box_dict['pred_boxes'].cpu().numpy()
            pred_labels = box_dict['pred_labels'].cpu().numpy()

            # Define an empty template dict to store the prediction information, 'pred_scores.shape[0]' means 'num_samples'
            pred_dict = get_template_prediction(pred_scores.shape[0])
            # If num_samples equals zero then return the empty dict
            if pred_scores.shape[0] == 0:
                return pred_dict

            # No calibration files

            pred_boxes_camera = box_utils.boxes3d_lidar_to_kitti_camera[pred_boxes]

            pred_dict['name'] = np.array(class_names)[pred_labels - 1]
            pred_dict['alpha'] = -np.arctan2(-pred_boxes[:, 1], pred_boxes[:, 0]) + pred_boxes_camera[:, 6]
            pred_dict['dimensions'] = pred_boxes_camera[:, 3:6]
            pred_dict['location'] = pred_boxes_camera[:, 0:3]
            pred_dict['rotation_y'] = pred_boxes_camera[:, 6]
            pred_dict['score'] = pred_scores
            pred_dict['boxes_lidar'] = pred_boxes

            return pred_dict

        annos = []
        for index, box_dict in enumerate(pred_dicts):
            frame_id = batch_dict['frame_id'][index]

            single_pred_dict = generate_single_sample_dict(index, box_dict)
            single_pred_dict['frame_id'] = frame_id
            annos.append(single_pred_dict)

            # Output pred results to Output-path in .txt file 
            if output_path is not None:
                cur_det_file = output_path / ('%s.txt' % frame_id)
                with open(cur_det_file, 'w') as f:
                    bbox = single_pred_dict['bbox']
                    loc = single_pred_dict['location']
                    dims = single_pred_dict['dimensions']  # lhw -> hwl: lidar -> camera

                    for idx in range(len(bbox)):
                        print('%s -1 -1 %.4f %.4f %.4f %.4f %.4f %.4f %.4f %.4f %.4f %.4f %.4f %.4f %.4f'
                            % (single_pred_dict['name'][idx], single_pred_dict['alpha'][idx],
                                bbox[idx][0], bbox[idx][1], bbox[idx][2], bbox[idx][3],
                                dims[idx][1], dims[idx][2], dims[idx][0], loc[idx][0],
                                loc[idx][1], loc[idx][2], single_pred_dict['rotation_y'][idx],
                                single_pred_dict['score'][idx]), file=f)
            return annos


    def __len__(self):
        if self._merge_all_iters_to_one_epoch:
            return len(self.sample_id_list) * self.total_epochs

        return len(self.custom_infos)


    def __getitem__(self, index):
        """
        Function:
            Read 'velodyne' folder as pointclouds
            Read 'label_2' folder as labels
            Return type 'dict'
        """
        if self._merge_all_iters_to_one_epoch:
            index = index % len(self.custom_infos)
        
        info = copy.deepcopy(self.custom_infos[index])

        sample_idx = info['point_cloud']['lidar_idx']
        get_item_list = self.dataset_cfg.get('GET_ITEM_LIST', ['points'])

        input_dict = {
            'frame_id': self.sample_id_list[index],
        }

        """
        Here infos was generated by get_infos
        """
        if 'annos' in info:
            annos = info['annos']
            annos = common_utils.drop_info_with_name(annos, name='DontCare')
            loc, dims, rots = annos['location'], annos['dimensions'], annos['rotation_y']
            gt_names = annos['name']
            gt_boxes_lidar = annos['gt_boxes_lidar']
        
        if 'points' in get_item_list:
            points = self.get_lidar(sample_idx, True)
            # import time
            # print(points.shape)
            # if points.shape[0] == 0:
            #     print("**********************************")
            #     print("sample_idx: ", sample_idx)
            #     time.sleep(999999)
            #     print("**********************************")
            # 000099, 000009
            input_dict['points'] = points
            input_dict.update({
                'gt_names': gt_names,
                'gt_boxes': gt_boxes_lidar
            })

        data_dict = self.prepare_data(data_dict=input_dict)
        return data_dict


def create_custom_infos(dataset_cfg, class_names, data_path, save_path, workers=4):
    dataset = CustomDataset(dataset_cfg=dataset_cfg, class_names=class_names, root_path=data_path, training=False)
    train_split, val_split = 'train', 'val'

    # No evaluation
    train_filename = save_path / ('custom_infos_%s.pkl' % train_split)
    val_filenmae = save_path / ('custom_infos%s.pkl' % val_split)
    trainval_filename = save_path / 'custom_infos_trainval.pkl'
    test_filename = save_path / 'custom_infos_test.pkl'

    print('------------------------Start to generate data infos------------------------')

    dataset.set_split(train_split)
    custom_infos_train = dataset.get_infos(num_workers=workers, has_label=True, count_inside_pts=True)
    with open(train_filename, 'wb') as f:
        pickle.dump(custom_infos_train, f)
    print('Custom info train file is save to %s' % train_filename)

    dataset.set_split('test')
    custom_infos_test = dataset.get_infos(num_workers=workers, has_label=False, count_inside_pts=False)
    with open(test_filename, 'wb') as f:
        pickle.dump(custom_infos_test, f)
    print('Custom info test file is saved to %s' % test_filename)

    print('------------------------Start create groundtruth database for data augmentation------------------------')
    dataset.set_split(train_split)
    # Input the 'custom_train_info.pkl' to generate gt_database
    dataset.create_groundtruth_database(train_filename, split=train_split)
    print('------------------------Data preparation done------------------------')

if __name__=='__main__':
    import sys
    if sys.argv.__len__() > 1 and sys.argv[1] == 'create_custom_infos':
        import yaml
        from pathlib import Path
        from easydict import EasyDict
        dataset_cfg = EasyDict(yaml.safe_load(open(sys.argv[2])))
        ROOT_DIR = (Path(__file__).resolve().parent / '../../../').resolve()
        create_custom_infos(
            dataset_cfg=dataset_cfg,
            class_names=['Car', 'Pedestrian', 'Cyclist'],
            data_path=ROOT_DIR / 'data' / 'custom',
            save_path=ROOT_DIR / 'data' / 'custom'
        )

新建tools/cfgs/dataset_configs/custom_dataset.yaml并修改

DATASET: 'CustomDataset'
DATA_PATH: '../data/custom'

# If this config file is modified then pcdet/models/detectors/detector3d_template.py:
# Detector3DTemplate::build_networks:model_info_dict needs to be modified.
POINT_CLOUD_RANGE: [-70.4, -40, -3, 70.4, 40, 1] # x=[-70.4, 70.4], y=[-40,40], z=[-3,1]

DATA_SPLIT: {
    'train': train,
    'test': val
}

INFO_PATH: {
    'train': [custom_infos_train.pkl],
    'test': [custom_infos_val.pkl],
}

GET_ITEM_LIST: ["points"]
FOV_POINTS_ONLY: True

POINT_FEATURE_ENCODING: {
    encoding_type: absolute_coordinates_encoding,
    used_feature_list: ['x', 'y', 'z', 'intensity'],
    src_feature_list: ['x', 'y', 'z', 'intensity'],
}

# Same to pv_rcnn[DATA_AUGMENTOR]
DATA_AUGMENTOR:
    DISABLE_AUG_LIST: ['placeholder']
    AUG_CONFIG_LIST:
        - NAME: gt_sampling
          # Notice that 'USE_ROAD_PLANE'
          USE_ROAD_PLANE: False
          DB_INFO_PATH:
              - custom_dbinfos_train.pkl # pcdet/datasets/augmentor/database_ampler.py:line 26
          PREPARE: {
             filter_by_min_points: ['Car:5', 'Pedestrian:5', 'Cyclist:5'],
             filter_by_difficulty: [-1],
          }

          SAMPLE_GROUPS: ['Car:20','Pedestrian:15', 'Cyclist:15']
          NUM_POINT_FEATURES: 4
          DATABASE_WITH_FAKELIDAR: False
          REMOVE_EXTRA_WIDTH: [0.0, 0.0, 0.0]
          LIMIT_WHOLE_SCENE: True

        - NAME: random_world_flip
          ALONG_AXIS_LIST: ['x']

        - NAME: random_world_rotation
          WORLD_ROT_ANGLE: [-0.78539816, 0.78539816]

        - NAME: random_world_scaling
          WORLD_SCALE_RANGE: [0.95, 1.05]

DATA_PROCESSOR:
    - NAME: mask_points_and_boxes_outside_range
      REMOVE_OUTSIDE_BOXES: True

    - NAME: shuffle_points
      SHUFFLE_ENABLED: {
        'train': True,
        'test': False
      }

    - NAME: transform_points_to_voxels
      VOXEL_SIZE: [0.05, 0.05, 0.1]
      MAX_POINTS_PER_VOXEL: 5
      MAX_NUMBER_OF_VOXELS: {
        'train': 16000,
        'test': 40000
      }

新建tools/cfgs/custom_models/pointrcnn.yaml并修改

CLASS_NAMES: ['Car']
# CLASS_NAMES: ['Car', 'Pedestrian', 'Cyclist']

DATA_CONFIG:
    _BASE_CONFIG_: /home/zonlin/CRLFnet/src/site_model/src/LidCamFusion/OpenPCDet/tools/cfgs/dataset_configs/custom_dataset.yaml
    _BASE_CONFIG_RT_: /home/zonlin/CRLFnet/src/site_model/src/LidCamFusion/OpenPCDet/tools/cfgs/dataset_configs/custom_dataset.yaml

    DATA_PROCESSOR:
        -   NAME: mask_points_and_boxes_outside_range
            REMOVE_OUTSIDE_BOXES: True

        -   NAME: sample_points
            NUM_POINTS: {
                'train': 16384,
                'test': 16384
            }

        -   NAME: shuffle_points
            SHUFFLE_ENABLED: {
                'train': True,
                'test': False
            }

MODEL:
    NAME: PointRCNN

    BACKBONE_3D:
        NAME: PointNet2MSG
        SA_CONFIG:
            NPOINTS: [4096, 1024, 256, 64]
            RADIUS: [[0.1, 0.5], [0.5, 1.0], [1.0, 2.0], [2.0, 4.0]]
            NSAMPLE: [[16, 32], [16, 32], [16, 32], [16, 32]]
            MLPS: [[[16, 16, 32], [32, 32, 64]],
                   [[64, 64, 128], [64, 96, 128]],
                   [[128, 196, 256], [128, 196, 256]],
                   [[256, 256, 512], [256, 384, 512]]]
        FP_MLPS: [[128, 128], [256, 256], [512, 512], [512, 512]]

    POINT_HEAD:
        NAME: PointHeadBox
        CLS_FC: [256, 256]
        REG_FC: [256, 256]
        CLASS_AGNOSTIC: False
        USE_POINT_FEATURES_BEFORE_FUSION: False
        TARGET_CONFIG:
            GT_EXTRA_WIDTH: [0.2, 0.2, 0.2]
            BOX_CODER: PointResidualCoder
            BOX_CODER_CONFIG: {
                'use_mean_size': True,
                'mean_size': [
                    [3.9, 1.6, 1.56],
                    [0.8, 0.6, 1.73],
                    [1.76, 0.6, 1.73]
                ]
            }

        LOSS_CONFIG:
            LOSS_REG: WeightedSmoothL1Loss
            LOSS_WEIGHTS: {
                'point_cls_weight': 1.0,
                'point_box_weight': 1.0,
                'code_weights': [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
            }

    ROI_HEAD:
        NAME: PointRCNNHead
        CLASS_AGNOSTIC: True

        ROI_POINT_POOL:
            POOL_EXTRA_WIDTH: [0.0, 0.0, 0.0]
            NUM_SAMPLED_POINTS: 512
            DEPTH_NORMALIZER: 70.0

        XYZ_UP_LAYER: [128, 128]
        CLS_FC: [256, 256]
        REG_FC: [256, 256]
        DP_RATIO: 0.0
        USE_BN: False

        SA_CONFIG:
            NPOINTS: [128, 32, -1]
            RADIUS: [0.2, 0.4, 100]
            NSAMPLE: [16, 16, 16]
            MLPS: [[128, 128, 128],
                   [128, 128, 256],
                   [256, 256, 512]]

        NMS_CONFIG:
            TRAIN:
                NMS_TYPE: nms_gpu
                MULTI_CLASSES_NMS: False
                NMS_PRE_MAXSIZE: 9000
                NMS_POST_MAXSIZE: 512
                NMS_THRESH: 0.8
            TEST:
                NMS_TYPE: nms_gpu
                MULTI_CLASSES_NMS: False
                NMS_PRE_MAXSIZE: 9000
                NMS_POST_MAXSIZE: 100
                NMS_THRESH: 0.85

        TARGET_CONFIG:
            BOX_CODER: ResidualCoder
            ROI_PER_IMAGE: 128
            FG_RATIO: 0.5

            SAMPLE_ROI_BY_EACH_CLASS: True
            CLS_SCORE_TYPE: cls

            CLS_FG_THRESH: 0.6
            CLS_BG_THRESH: 0.45
            CLS_BG_THRESH_LO: 0.1
            HARD_BG_RATIO: 0.8

            REG_FG_THRESH: 0.55

        LOSS_CONFIG:
            CLS_LOSS: BinaryCrossEntropy
            REG_LOSS: smooth-l1
            CORNER_LOSS_REGULARIZATION: True
            LOSS_WEIGHTS: {
                'rcnn_cls_weight': 1.0,
                'rcnn_reg_weight': 1.0,
                'rcnn_corner_weight': 1.0,
                'code_weights': [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
            }

    POST_PROCESSING:
        RECALL_THRESH_LIST: [0.3, 0.5, 0.7]
        SCORE_THRESH: 0.1
        OUTPUT_RAW_SCORE: False

        EVAL_METRIC: kitti

        NMS_CONFIG:
            MULTI_CLASSES_NMS: False
            NMS_TYPE: nms_gpu
            NMS_THRESH: 0.1
            NMS_PRE_MAXSIZE: 4096
            NMS_POST_MAXSIZE: 500


OPTIMIZATION:
    BATCH_SIZE_PER_GPU: 2
    NUM_EPOCHS: 80

    OPTIMIZER: adam_onecycle
    LR: 0.01
    WEIGHT_DECAY: 0.01
    MOMENTUM: 0.9

    MOMS: [0.95, 0.85]
    PCT_START: 0.4
    DIV_FACTOR: 10
    DECAY_STEP_LIST: [35, 45]
    LR_DECAY: 0.1
    LR_CLIP: 0.0000001

    LR_WARMUP: False
    WARMUP_EPOCH: 1

    GRAD_NORM_CLIP: 10

其他调整事项

需要对以上文件中的类别信息，数据集路径，点云范围POINT_CLOUD_RANGE进行更改
在 pcdet/datasets/init.py文件，增加以下代码

from .custom.custom_dataset import CustomDataset
# 在__all__ = 中增加
'CustomDataset': CustomDataset

完成以上就可以开始对数据集进行预处理和训练了

数据集预处理

python -m pcdet.datasets.custom.custom_dataset create_custom_infos tools/cfgs/dataset_configs/custom_dataset.yaml

同时在gt_database文件夹下生成的.bin文件，data/custom文件夹结构变为如下：

custom
├── ImageSets
│   ├── test.txt
│   ├── train.txt
├── testing
│   ├── velodyne
├── training
│   ├── label_2
│   ├── velodyne
├── gt_database
│   ├── xxxxx.bin
├── custom_infos_train.pkl
├── custom_infos_val.pkl
├── custom_dbinfos_train.pkl

数据集训练

python tools/train.py --cfg_file tools/cfgs/custom_models/pointrcnn.yaml --batch_size=2 --epochs=300

可视化测试

cd到tools文件夹下，运行：

python demo.py --cfg_file cfgs/custom_models/pointrcnn.yaml  --data_path ../data/custom/testinging/velodyne/ --ckpt ../output/custom_models/pointrcnn/default/ckpt/checkpoint_epoch_300.pth

此处根据自己的文件路径进行修改，推理效果如下（笔者标注50多张闸口船舶的点云数据）：

看起来效果还是挺不错。

获取尺寸

OpenPCDet平台下根据kitti格式推理得到的bbox的dx、dy、dz就是约等于实际的物体的尺寸。

对于我们的点云数据而言，上述数据对应船的高宽长。（这里不理解的可以去看下OpenPCDet的坐标定义）

---

四、总结

至此，基于OpenPCDet平台的自定义数据集的训练基本完成了，这里要特别感谢下树和猫，对于自定义数据集的训练我们交流了很多，之前他是通过我写的yolov5系列文章关注的我，现在我通过OpenPCDet 训练自己的数据集系列关注了他，着实让我感觉到了技术分享是一个圈。

参考文档：
https://blog.csdn.net/m0_68312479/article/details/126201450
https://blog.csdn.net/jin15203846657/article/details/122949271
https://blog.csdn.net/hihui1231/article/details/124903276
https://github.com/OrangeSodahub/CRLFnet/tree/master/src/site_model/src/LidCamFusion/OpenPCDet
https://blog.csdn.net/weixin_43464623/article/details/116718451

如果阅读本文对你有用，欢迎一键三连呀！！！
2022年10月24日11:12:53