引導式測驗 · 硬 等級
RANSAC穩健模型擬合測驗
探索 RANSAC 採樣、共識和殘差閾值、簡併幾何、迭代預算以及穩健模型擬合的限制。
問題 1 的 8
A photo matcher includes several wrong point pairs. What does one RANSAC trial do?
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。
- A photo matcher includes several wrong point pairs. What does one RANSAC trial do?
- Fit least squares to all matches before checking outliers
- Choose only the lowest descriptor-distance pair
- Average point coordinates without proposing geometry
- Fit a candidate from a small sample and score how many observations agree
- Why do low inlier fractions usually require more RANSAC trials?
- The minimal sample becomes smaller as outliers increase
- An all-inlier minimal sample becomes less likely
- The residual threshold automatically removes bad samples
- Each trial is guaranteed to include at least one inlier
- A general homography is proposed from four point pairs that lie on one line. What should an implementation do?
- Reject the degenerate sample and try another
- Accept four pairs regardless of geometric arrangement
- Set a zero residual threshold for the collinear pairs
- Use the same degenerate sample for every later trial
- If the inlier residual threshold is set too loose, what risk grows?
- Legitimate noisy points are always counted as inliers
- The selected model necessarily has fewer parameters
- Wrong correspondences may be counted as support for a misleading model
- Every point becomes an inlier independent of its residual
- Why refit the chosen model using its consensus set?
- The refit guarantees a globally correct scene model
- It changes the estimated homography into a different model family
- It makes residual-threshold selection unnecessary
- More supported observations can improve parameter estimates beyond the minimal sample
- A result has many inliers, but all lie in one tiny image corner. What concern remains?
- High inlier count guarantees broad image coverage
- The matches may poorly constrain alignment elsewhere in the image
- A tight cluster constrains distant image regions equally well
- Spatial concentration rules out a second plane
- Which quantity should set a pixel-based homography inlier threshold?
- The raw number of feature matches before geometric fitting
- The descriptor length used by the feature matcher
- A meaningful pixel reprojection error for the expected localization noise
- A confidence target without any measurement scale
- Which problem is RANSAC least able to solve by itself?
- A model family that does not describe the actual scene
- A modest fraction of false feature matches
- A few isolated point-cloud outliers
- One incorrectly paired image feature