Bài kiểm tra có liên kết hướng dẫn · Cứng Cấp độ
Bài kiểm tra ước tính tư thế vật thể 6D
Kiểm tra khả năng dịch và xoay 6D, khung hình tương đối với máy ảnh, sự tương ứng với mô hình, tính điểm nhận biết đối xứng và cách sử dụng tư thế trong robot và AR.
Câu hỏi 1 của 8
A detector gives a tight 2D box around a tool. Which information is still needed for a gripper to approach it?
Xem trước mọi câu hỏi8 câu hỏi
Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.
- 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
- 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
- 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
- An unmarked cylinder looks the same after rotation around its axis. How should evaluation handle this?
- 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
- What can RGB-D data add to an RGB-only pose estimate?
- 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
- 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
- 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
- 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