随记(之后再整理)

单纯的想复制代码…

class SemanticDataset:
    def __init__(
        self, num_points_per_sample, split, use_color, box_size_x, box_size_y, path
    ):
        
        # Dataset parameters
        self.num_points_per_sample = num_points_per_sample
        self.split = split
        self.use_color = use_color
        self.box_size_x = box_size_x
        self.box_size_y = box_size_y
        self.num_classes = 9
        self.path = path
        self.labels_names = [
            "unlabeled",
            "man-made terrain",
            "natural terrain",
            "high vegetation",
            "low vegetation",
            "buildings",
            "hard scape",
            "scanning artifact",
            "cars",
        ]

        # Get file_prefixes
        file_prefixes = map_name_to_file_prefixes[self.split]
        print("Dataset split:", self.split)
        print("Loading file_prefixes:", file_prefixes)

        # Load files
        self.list_file_data = []
        for file_prefix in file_prefixes:
            file_path_without_ext = os.path.join(self.path, file_prefix)
            file_data = SemanticFileData(
                file_path_without_ext=file_path_without_ext,
                has_label=self.split != "test",
                use_color=self.use_color,
                box_size_x=self.box_size_x,
                box_size_y=self.box_size_y,
            )
            self.list_file_data.append(file_data)

        # Pre-compute the probability of picking a scene
        self.num_scenes = len(self.list_file_data)
        self.scene_probas = [
            len(fd.points) / self.get_total_num_points() for fd in self.list_file_data
        ]

        # Pre-compute the points weights if it is a training set
        if self.split == "train" or self.split == "train_full":
            # First, compute the histogram of each labels
            label_weights = np.zeros(9)
            for labels in [fd.labels for fd in self.list_file_data]:
                tmp, _ = np.histogram(labels, range(10))
                label_weights += tmp

            # Then, a heuristic gives the weights
            # 1 / log(1.2 + probability of occurrence)
            label_weights = label_weights.astype(np.float32)
            label_weights = label_weights / np.sum(label_weights)
            self.label_weights = 1 / np.log(1.2 + label_weights)
        else:
            self.label_weights = np.zeros(9)

读取数据集，for循环，对每个数据集都做以下提取操作

class SemanticFileData:
    def __init__(
        self, file_path_without_ext, has_label, use_color, box_size_x, box_size_y
    ):
        """
        Loads file data
        """
        self.file_path_without_ext = file_path_without_ext
        self.box_size_x = box_size_x
        self.box_size_y = box_size_y

        # Load points
        pcd = open3d.read_point_cloud(file_path_without_ext + ".pcd")
        self.points = np.asarray(pcd.points)

        # Load label. In pure test set, fill with zeros.
        if has_label:
            self.labels = load_labels(file_path_without_ext + ".labels")
        else:
            self.labels = np.zeros(len(self.points)).astype(bool)

        # Load colors. If not use_color, fill with zeros.
        if use_color:
            self.colors = np.asarray(pcd.colors)
        else:
            self.colors = np.zeros_like(self.points)

        # Sort according to x to speed up computation of boxes and z-boxes
        sort_idx = np.argsort(self.points[:, 0])
        self.points = self.points[sort_idx]
        self.labels = self.labels[sort_idx]
        self.colors = self.colors[sort_idx]

数据预处理，将坐标，标签，颜色，权重全部提取出来