Design a YAML-Configured Point-Cloud Transformation Pipeline

Design a YAML-Configured Point-Cloud Transformation Pipeline

In autonomous driving perception systems, raw LiDAR point clouds undergo sequential preprocessing (e.g., coordinate transformation, range filtering, voxel grid downsampling) before being fed into 3D perception models. These transformation pipelines must be modular, declarative (configured via YAML/dicts), validated ahead of runtime, and track sensor metadata across transformations.

Implement an extensible point-cloud processing pipeline in Python consisting of:

1. PointCloud Data Structure: A data container holding:

   • points: A NumPy array of shape (N, D) where D >= 3 (e.g., x, y, z, intensity).
   • metadata: A dictionary holding tracking info (e.g., timestamp, frame_id, and a list of applied_transforms).

2. Transform Registry & Base Interface:

   • An abstract base class BaseTransform with an apply(pc: PointCloud) -> PointCloud method.
   • A decorator or registry mechanism @register_transform(name) that registers transforms by string identifier.
   • Each transform must validate its initialization arguments against expected types and ranges.

3. Core Transforms:

   • CropBoundingBox: Filters points within [min_x, min_y, min_z] and [max_x, max_y, max_z] inclusive. Appends {"name": "CropBoundingBox", "remaining_points": count} to metadata["applied_transforms"].
   • RigidTransform: Applies a 4x4 homogeneous transformation matrix to (x, y, z) coordinates (preserving any extra features like intensity). Updates metadata["frame_id"] to the transform's target frame.

4. TransformPipeline:

   • Initialized with a list of transform configurations (parsed from YAML as a list of dicts: [{"name": "...", "params": {...}}, ...]).
   • Validates the entire pipeline configuration upon initialization (raises ValueError for unknown transform names or invalid parameter types/bounds).
   • Implements run(pc: PointCloud) -> PointCloud, applying transforms sequentially.

Interface

from abc import ABC, abstractmethod
from typing import Any, Dict, List
import numpy as np

class PointCloud:
    def __init__(self, points: np.ndarray, metadata: Dict[str, Any] = None):
        self.points = points
        self.metadata = metadata if metadata is not None else {}
        if "applied_transforms" not in self.metadata:
            self.metadata["applied_transforms"] = []

class TransformRegistry:
    # Implement registry mechanism
    pass

class BaseTransform(ABC):
    @abstractmethod
    def apply(self, pc: PointCloud) -> PointCloud:
        pass

class TransformPipeline:
    def __init__(self, config: List[Dict[str, Any]]):
        pass

    def run(self, pc: PointCloud) -> PointCloud:
        pass

Example

config = [
    {
        "name": "CropBoundingBox",
        "params": {
            "min_bound": [-10.0, -10.0, -2.0],
            "max_bound": [10.0, 10.0, 2.0]
        }
    },
    {
        "name": "RigidTransform",
        "params": {
            "target_frame": "vehicle",
            "matrix": np.eye(4).tolist()
        }
    }
]

pipeline = TransformPipeline(config)
points = np.array([
    [0.0, 0.0, 0.0, 0.8],
    [20.0, 0.0, 0.0, 0.5],
    [-5.0, 5.0, 1.0, 0.2]
])
pc = PointCloud(points=points, metadata={"frame_id": "lidar_top"})

result = pipeline.run(pc)
print(result.points.shape)  # (2, 4) -> (20, 0, 0) was cropped out
print(result.metadata["frame_id"])  # "vehicle"
print(len(result.metadata["applied_transforms"]))  # 2

Constraints

• points array has shape (N, D) where N >= 0 and D >= 3.
• All transforms must return a new or modified PointCloud without mutating unmodified attributes destructively.
• If an unknown transform is passed to TransformPipeline, raise ValueError immediately at construction time, not during run.

Edge Cases

• Empty point cloud (N = 0): Transforms should execute cleanly without throwing dimension errors.
• Point cloud with additional feature dimensions (D > 3, e.g., intensity, elongation, ring ID): RigidTransform must transform only the first 3 coordinates while keeping additional channels unchanged.