bev-project/mmdet3d/datasets/pipelines/loading_vector.py

"""
矢量地图数据加载Pipeline
"""

import pickle
import numpy as np
from mmdet.datasets import PIPELINES


@PIPELINES.register_module()
class LoadVectorMap:
    """
    加载矢量地图标注
    
    从预处理的pkl文件中加载矢量地图数据
    
    Args:
        vector_map_file: 矢量地图数据文件路径
        num_vec_classes: 矢量类别数
        max_num_vectors: 每个样本最多的矢量数
        num_points_per_vec: 每个矢量的点数
    """
    
    def __init__(
        self,
        vector_map_file='data/nuscenes/vector_maps_bevfusion.pkl',
        num_vec_classes=3,
        max_num_vectors=50,
        num_points_per_vec=20,
    ):
        self.num_vec_classes = num_vec_classes
        self.max_num_vectors = max_num_vectors
        self.num_points_per_vec = num_points_per_vec
        
        # 加载矢量地图数据
        print(f"加载矢量地图数据: {vector_map_file}")
        with open(vector_map_file, 'rb') as f:
            self.vector_maps = pickle.load(f)
        
        print(f"✅ 加载了 {len(self.vector_maps)} 个样本的矢量地图")
        
        # 类别映射
        self.class_map = {
            'divider': 0,
            'boundary': 1,
            'ped_crossing': 2,
        }
    
    def __call__(self, results):
        """
        加载当前样本的矢量地图
        
        Args:
            results: dict, 包含 'token' 或 'sample_idx'
        
        Returns:
            results: 添加 'gt_vectors_labels' 和 'gt_vectors_points'
        """
        # 获取样本标识
        sample_token = results.get('token', results.get('sample_idx', None))
        
        if sample_token is None:
            print("警告: 无法获取sample_token，使用空矢量地图")
            vectors_raw = {'divider': [], 'boundary': [], 'ped_crossing': []}
        else:
            vectors_raw = self.vector_maps.get(sample_token, 
                                              {'divider': [], 'boundary': [], 'ped_crossing': []})
        
        # 转换为训练格式
        gt_vectors = []
        
        # Divider (class 0)
        for vec in vectors_raw.get('divider', []):
            points = np.array(vec['points'], dtype=np.float32)
            if len(points) == self.num_points_per_vec:
                gt_vectors.append({
                    'class': 0,
                    'points': points,
                })
        
        # Boundary (class 1)
        for vec in vectors_raw.get('boundary', []):
            points = np.array(vec['points'], dtype=np.float32)
            if len(points) == self.num_points_per_vec:
                gt_vectors.append({
                    'class': 1,
                    'points': points,
                })
        
        # Ped crossing (class 2)
        for vec in vectors_raw.get('ped_crossing', []):
            points = np.array(vec['points'], dtype=np.float32)
            if len(points) == self.num_points_per_vec:
                gt_vectors.append({
                    'class': 2,
                    'points': points,
                })
        
        # 限制数量（截断）
        if len(gt_vectors) > self.max_num_vectors:
            # 随机采样或按顺序截断
            gt_vectors = gt_vectors[:self.max_num_vectors]
        
        # Padding到固定数量
        while len(gt_vectors) < self.max_num_vectors:
            gt_vectors.append({
                'class': -1,  # 无效类别（padding）
                'points': np.zeros((self.num_points_per_vec, 2), dtype=np.float32),
            })
        
        # 转为numpy数组
        gt_labels = np.array([v['class'] for v in gt_vectors], dtype=np.int64)
        gt_points = np.array([v['points'] for v in gt_vectors], dtype=np.float32)
        
        # 添加到results
        results['gt_vectors_labels'] = gt_labels
        results['gt_vectors_points'] = gt_points
        
        return results
    
    def __repr__(self):
        repr_str = self.__class__.__name__
        repr_str += f'(num_vec_classes={self.num_vec_classes}, '
        repr_str += f'max_num_vectors={self.max_num_vectors}, '
        repr_str += f'num_points_per_vec={self.num_points_per_vec})'
        return repr_str