Викторина, связанная с руководством · Жесткий Уровень

Тест на оценку позы 6D-объекта

Тестируйте перемещение и вращение 6D, кадры относительно камеры, соответствие моделей, оценку с учетом симметрии и использование поз в роботах и дополненной реальности.

Связанные направляющие пути6d Оценка позы объекта
Вопрос 1 из 8

A detector gives a tight 2D box around a tool. Which information is still needed for a gripper to approach it?

Предварительный просмотр каждого вопроса8 вопросы

Ответы и пояснения появляются по мере прохождения теста.

  1. A detector gives a tight 2D box around a tool. Which information is still needed for a gripper to approach it?
    • Only the class label and detector confidence
    • Only a 2D mask around the visible pixels
    • Only distance to the nearest surface without orientation
    • The tool’s 3D position and orientation in a known frame
  2. In a rigid 6D pose, what do the six degrees describe?
    • Three translation and three rotational degrees
    • Three translations and three independent scale factors
    • Six articulation angles for a deformable object
    • Three colors plus three depth channels
  3. Why should camera intrinsics be known when projecting a 3D model onto an image?
    • They supply the object’s rigid transform relative to the camera
    • They directly remove occlusion from the image
    • They map camera-frame points to image coordinates
    • They determine the object’s real-world identity
  4. Незначимый цилиндр выглядит одинаково после вращения вокруг своей оси. Как следует обрабатывать оценку в этом случае?
    • Penalize every angle other than one stored annotation
    • Treat visually or geometrically equivalent rotations as equivalent where the task allows
    • Ignore all translation errors for symmetric objects
    • Assume a unique orientation from geometry that is unchanged
  5. Что может добавить данные RGB‑D к оценке позы только по RGB?
    • Measured range cues that may help constrain object geometry and translation
    • An exact pose on every reflective surface
    • A built-in 3D model for every unknown object
    • A semantic identity label for each pixel
  6. A pose method trained on synthetic renders is tested on real camera images. Which risk deserves measurement?
    • Real camera appearance is guaranteed to match rendering
    • Synthetic training removes the need for calibration
    • A domain gap in appearance and sensor characteristics
    • Rigid pose has extra degrees only in real scenes
  7. Why might box-overlap accuracy be insufficient for robotic pose evaluation?
    • Good 2D overlap guarantees correct 3D rotation
    • Two boxes can overlap while orientations or 3D translations differ enough to spoil a grasp
    • The box alone specifies a collision-free grasp
    • Pose error matters only to image compression
  8. What assumption limits a known-model 6D pose method on an unfamiliar object?
    • A known model can be recovered from the class name alone
    • RGB input cannot be used with known models
    • Object coordinates cease to exist without depth
    • It may lack the correct 3D shape and scale needed for matching