引導式測驗 · 硬 等級
單張照片測驗的肖像動畫
測試基於關鍵點的動作傳輸、遮蔽處理以及 SadTalker 和 LivePortrait 設計選擇,以對一張照片中的臉部進行動畫處理。
問題 1 的 8
First Order Motion Model 如何取得其關鍵點?
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。
- First Order Motion Model 如何取得其關鍵點?
- From manually labeled facial landmarks
- From a 3D scan of the person
- It learns them without labels during training
- From the audio track
- 遮擋圖(occlusion map)的用途是什麼?
- To mark regions the source image cannot supply, so they must be painted in
- To blur the background on purpose
- To detect the speaker's language
- To choose the output frame rate
- 相對動作轉移(relative motion transfer)有什麼好處?
- It doubles the frame rate
- The source keeps its own face shape instead of taking on the driver's
- It removes the need for a driving video
- It makes the head turn farther
- SadTalker 從音訊中預測什麼?
- Raw pixel values of the mouth
- A new background
- 文字記錄
- 3D Morphable Model coefficients for expression and head pose
- LivePortrait 的拼接模組(stitching module)有什麼作用?
- It pastes the animated face back into the full image without misaligned shoulders
- It merges two different people's faces
- It stitches audio clips together
- It converts video to GIF
- Why do large head turns usually fail in single-photo animation?
- The GPU runs out of memory
- Head turns are disabled by default
- The photo has no information about unseen regions like the side of the head
- Audio cannot describe rotation
- Why do keypoint methods warp feature maps rather than raw pixels?
- Pixels cannot be moved
- It lets the decoder repair stretching and fill unseen regions
- Feature maps are always higher resolution
- It removes the need for keypoints
- In implicit 3D keypoint methods, what does the driving frame contribute?
- The canonical keypoints encoding identity
- The background texture
- The skin color
- Head rotation, translation and expression deformations