3D Detection Evaluation

evaluate_3d_detections.py

import fiftyone as fo
from fiftyone import ViewField as F

import numpy as np
from numpy.random import random as npr
import open3d as o3d

samples = []


for j in range(50):
    gt_detections = []
    pred_detections = []
    point_cloud = []
    
    for _ in range(10):
        dimensions = np.random.uniform([1, 1, 1], [3, 3, 3])
        location = np.random.uniform([-10, -10, 0], [10, 10, 10])
        rotation = np.random.uniform(-np.pi, np.pi, size=3)
        conf = npr()

        gt_detection = fo.Detection(
            label = "dog",
            dimensions=list(dimensions),
            location=list(location),
            rotation=list(rotation),
        )
        gt_detections.append(gt_detection)

        pred_detection = fo.Detection(
            label = "dog",
            confidence=conf,
            dimensions=list(dimensions + 0.3 * npr()),
            location=list(location + 0.3 * npr()),
            rotation=list(rotation + 0.3 * npr()),
        )
        
        if npr() > 0.2:
            pred_detections.append(pred_detection)

        R = o3d.geometry.get_rotation_matrix_from_xyz(rotation)
        points = np.random.uniform(-dimensions / 2, dimensions / 2, size=(1000, 3))
        points = points @ R.T + location[np.newaxis, :]
        point_cloud.extend(points)

    pc = o3d.geometry.PointCloud()
    pc.points = o3d.utility.Vector3dVector(np.array(point_cloud))
    o3d.io.write_point_cloud(f"/tmp/toy{j}.pcd", pc)

    sample = fo.Sample(
            filepath=f"/tmp/toy{j}.pcd",
            ground_truth=fo.Detections(detections=gt_detections),
            predictions=fo.Detections(detections=pred_detections),  
        )
    samples.append(sample)

dataset = fo.Dataset("3d-eval", persistent = True)
dataset.add_samples(samples)

## set in-app visualization configuration
dataset.app_config.plugins["3d"] = {
    "defaultCameraPosition": {"x": 0, "y": 0, "z": 20}
}

fo.launch_app(dataset)