引導式測驗 · 硬 等級

單眼深度估計測驗

評估單目深度模糊性、自監督視圖合成、相對尺度和特定版本的深度模型證據。

相關引導路徑單目深度估計
問題 1 的 8

是什麼讓深度估計成為「單目」?

預覽每個問題8 問題

當您參加測驗時,就會出現答案和解釋。

  1. 是什麼讓深度估計成為「單目」?
    • 它從單一影像或相機推斷深度,而不是立體或感測器
    • 它使用兩個像人眼一樣間隔的攝像頭
    • 它僅適用於黑白影像
    • 它需要連接到相機的雷射雷達掃描儀
  2. What does a monocular depth estimator predict for an image?
    • A stereo disparity field that requires two camera views
    • A per-pixel depth or relative-depth map from one camera view
    • A categorical object-label map with no distance estimate
    • A dense motion-vector field between two video frames
  3. What data recipe does the Depth Anything V2 paper report?
    • Train only on real labeled images and discard synthetic data
    • Train on unlabeled audio and convert speech to depth maps
    • Train a DINOv2-G teacher on synthetic labeled images, then use its pseudo-labels on real images to train students
    • Use stereo disparity as labels for every image at inference time
  4. 單目網路不直接依賴哪一種線索來獲得深度?
    • 物體的相對大小
    • 紋理漸層和透視
    • 遮蔽(哪個物體阻擋另一個物體)
    • 兩台同步攝影機之間的立體視差
  5. What does Monodepth2 use to reduce the effect of moving objects in self-supervised training?
    • An auto-masking procedure that excludes pixels inconsistent with camera motion
    • A learned text tokenizer
    • A region-proposal network
    • A fixed depth value for every object
  6. What new evaluation benchmark does the Depth Anything V2 paper introduce?
    • A benchmark of 2,000 full metric depth maps from LiDAR only
    • DA-2K, with sparse relative-depth pairs annotated across 1,000 images
    • A corpus of six public datasets and randomly captured photos
    • A camera-calibration set containing no depth comparisons
  7. How can self-supervised monocular depth training obtain a learning signal?
    • By reading depth labels from every pixel
    • By converting image colors directly to meters
    • By reconstructing one view from adjacent video frames and penalizing photometric mismatch
    • By using only language captions
  8. What does a relative-depth output indicate?
    • Exact meters for every pixel
    • Camera GPS position
    • Object identity only
    • Ordering or relative distance, not necessarily calibrated metric distance