The data association problem in SLAM

WHAT THIS TESTS The interviewer wants to know if you understand the hidden assumption behind every SLAM optimization: that observations are correctly matched to landmarks. Get associations wrong and the math optimizes the wrong constraints.

A GOOD ANSWER COVERS Definition: data association is determining which current measurement corresponds to which existing map element or previous observation, both frame to frame and during loop closure. It is critical because bundle adjustment and filters assume correspondences are correct; even a few false matches bias poses and landmarks, and errors compound over time. Techniques include feature descriptor matching like ORB with ratio tests, geometric verification with RANSAC on the fundamental or essential matrix, motion-model or covariance gating to limit the search region, and appearance-based bag-of-words for place recognition during loop closure, plus joint compatibility tests for stronger consistency.

COMMON WRONG ANSWERS Calling it a simple nearest-neighbor lookup with no outlier handling. Or conflating it only with loop closure, when it also happens every frame. Or ignoring perceptual aliasing, where repetitive structures cause confident wrong matches.

LIKELY FOLLOW-UPS How does RANSAC reject outlier matches. What is perceptual aliasing and how does it cause false loop closures. How does gating reduce the search space. Why are wrong loop closures especially catastrophic.

ONE CONCRETE EXAMPLE In a hallway with identical doors, the place-recognition module proposes a loop closure between two different doors that look alike, a case of perceptual aliasing. Without geometric verification, global optimization would trust the false constraint and fold the map incorrectly, ruining the trajectory. RANSAC on the feature correspondences fits a geometric model and finds the proposed match is inconsistent with the implied camera motion, so it rejects the loop and prevents a corrupted map. The same machinery, applied every frame with descriptor matching and motion-model gating, keeps frame-to-frame associations reliable too.