引導式測驗 · 硬 等級
AI 視訊生成測驗中的相機控制
檢查變焦與移動車、Plücker 相機嵌入、相機數據稀缺的原因以及如何測量 AI 影片中的軌跡準確性。
問題 1 的 8
行動攝影車和變焦鏡頭在視覺上有何不同?
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。
- 行動攝影車和變焦鏡頭在視覺上有何不同?
- A zoom always moves left to right
- A dolly moves the camera and creates parallax, while a zoom changes focal length without parallax
- A dolly changes focal length only
- 它們是相同的
- What does a Plücker embedding encode for each pixel?
- Its color value
- Its depth in meters
- The object label at that pixel
- The viewing ray from the camera pose, as (o × d, d)
- 為什麼很難從典型的訓練影片中學習精確的攝影機控制?
- Most videos have no camera labels, and captions rarely describe camera motion accurately
- Videos are always filmed on tripods
- Camera motion cannot be seen in video
- Training videos have no frames
- 像 RealEstate10K 這樣的帶有姿勢註釋的資料集有什麼限制?
- They contain no camera poses
- They are audio-only
- They are narrow in domain and mostly show static scenes
- They only include animated films
- 如何衡量產生的影片是否遵循要求的攝影機路徑?
- Count the number of frames
- Run structure-from-motion on the output and compare the recovered path using rotation and translation errors
- Check the file size
- Ask the model to describe its camera
- What does scale ambiguity mean for camera control?
- Videos cannot be resized
- Zooming always fails
- Pixels have no color scale
- Monocular video has no absolute scale, so distances like 'two meters' are undefined without normalization
- How does CameraCtrl add camera conditioning to a video model?
- It trains an adapter on pose-annotated clips while the base model stays frozen
- It retrains the entire model from scratch
- It edits frames after generation
- It uses text prompts only
- Which prompting practice gives more reliable camera moves in prompt-based tools?
- Request several moves at once
- Describe only the subject
- Ask for one move per shot, state its speed, and keep subject motion simple
- Avoid all camera words