SLAM, Localization, and Mapping Job-Hunting Notes

I learned a great deal from others’ job-hunting experience and would like to share my own notes and suggestions for SLAM positions.

Timing

My job search lasted a little more than two months, from late March to early June 2019, near the spring-recruitment period. Searching during spring recruitment was difficult: many companies had not released graduate positions and submitting résumés was hard. Try to apply during autumn recruitment, and apply early—the earlier, the more opportunities.

Résumé

Fit is the most important thing in a résumé. Tailor it to the job description, putting relevant and important information near the top where it is most visible.

Referrals are preferable. None of the résumés I submitted without a referral led to an interview.

When a referral reaches a department lead, attach a short project introduction in addition to the résumé. Images and text can explain the project clearly and create a stronger impression.

Interviews

A résumé only opens the door to an interview; the interview is an opportunity to demonstrate more. In-person interviews are useful because they support richer demonstrations and easier communication. Prepare a presentation that tells a coherent story about your work; bring diagrams, images, or video if useful.

Confidence matters. Keep going.

Interview questions

My target companies worked on autonomous driving or robotics. Interviews varied: some first asked candidates to introduce projects and then tested fundamental SLAM and algorithm knowledge; others drilled into the résumé directly. The following are questions I encountered.

1. SLAM questions

1. What is loop closure? What methods are common? Which did you use, and what was novel?
2. Explain Gauss–Newton and LM.
3. Are you familiar with Ceres? Explain it.
4. Describe the (extended) Kalman filter and particle filter. What problems did you encounter using Kalman filtering?
5. Have you used sensors beyond vision, such as GPS or LiDAR?
6. What are tightly and loosely coupled systems? What are their pros and cons?
7. What differs between indoor SLAM and autonomous-driving SLAM?
8. What methods construct map points?
9. Given matched 2D features across consecutive frames with possible bad matches, how would you construct 3D points?
10. What metric does RANSAC use to select its best model?
11. What robust-estimation methods exist besides RANSAC?
12. What robust kernels are available?
13. How are 3D map points stored and represented?
14. In bundle adjustment with $m$ cameras and $n$ points, why can camera poses approach ground truth much faster than map points during simulation?
15. How does $\\lambda$ change in LM?
16. How can 3D poses be represented?
17. What is the relationship between Lie groups and Lie algebras?
18. Derive $\\frac{\\partial R_1R_2}{\\partial R_1}$.
19. Given an $H$-by-$W$ image or matrix, should elements be visited by row or column? How does this relate to cache?
20. Write the monocular-camera projection and distortion models.
21. Describe a SLAM algorithm you know—LiDAR or visual—and its advantages and limitations.
22. Read Maplab and design a map-update workflow for an indoor service robot.
23. Write code to deskew LiDAR data from a platform rotating at constant speed.
24. Given two matched 3D point sets, calculate the relative pose transformation—known-correspondence ICP—and write code.
25. Why does ORB-SLAM initialization calculate both homography $H$ and fundamental matrix $F$?
26. Explain the Dog-Leg algorithm.
27. What are marginalization, the First-Estimate Jacobian, consistency, and observability?
28. What are the advantages and limitations of VINS-Mono?
29. What is the essential difference, contribution, or novelty of your work?
30. What are the essential and fundamental matrices?
31. Given continuous pose-known frames, how would you measure lane-line coordinates in the world frame?
32. How would you accurately localize with noisy GPS?
33. How would you calibrate IMU–camera extrinsics?
34. Given GPS and inertial-navigation errors, how would you obtain a centimeter-level map?
35. What tricks are useful when calculating $H$ and $F$? This is really asking about normalization.
36. Extract planes from a point cloud.
37. Given an image and the camera-to-ground relationship, compute a bird’s-eye view.
38. Write the formula for bilinear interpolation.
39. What differs between RGB-D SLAM and RGB SLAM?
40. Design a localization system for a delivery robot travelling 3 km from a supermarket to a residential area: include sensors, algorithm flow, and pseudocode.
41. What is ORB? How is its rotation invariance achieved, and how is BRIEF extracted?
42. Write the formula and workflow for image undistortion.
43. How are features extracted and distributed uniformly in ORB-SLAM?

2. Algorithms, data structures, and C++

1. What structure is the ORB-SLAM covisibility graph, and how is it stored?
2. Write a quadtree.
3. Traverse a binary tree without recursion.
4. Find the largest connected component.
5. How can multithreading be implemented?
6. Describe std::vector and how it grows dynamically or shrinks capacity.
7. Is Eigen row-major or column-major?
8. What is the difference between unordered and ordered containers?
9. How is Mat destructed?
10. Explain smart pointers: shared_ptr and unique_ptr.
11. What is a virtual function?
12. How can ordinary pointers enforce unique ownership of a memory block?
13. What is C++ RTTI?
14. How does C++ implement polymorphism?
15. When does a vector iterator become invalid?
16. Reconstruct a binary tree.
17. Write a CMakeLists.txt and gcc command.
18. Given a graph and pairs of nodes, find a path between each pair.
19. Implement a sparse-matrix data structure and sparse-matrix addition.
20. In a checkerboard with a light at every corner, pressing a button toggles four neighboring lights. Given a random initial state, how can all lights be turned off?
21. Implement binary search.
22. Sort an array and discuss sorting algorithms.
23. Given two 2D line segments, determine whether they intersect.
24. Implement quicksort and reverse a linked list.
25. Given two ascending sorted arrays of sizes $n$ and $m$, find the $k$th smallest number.

Preparation

Many experienced practitioners already provide strong preparation advice and resources [1–6]. Foundations matter more in today’s job market. If possible, implement core SLAM algorithms yourself, including PnP, ICP, and BA, and practice basic algorithm problems.

I wish everyone the best in finding an ideal role.

References

1. SLAM job-hunting handbook
2. SLAM and 3D-vision written-test and interview questions
3. SLAM and 3D-vision job-hunting experience
4. SLAM summer-internship job-hunting notes
5. CS PhD SLAM / autonomous-driving job-hunting summary
6. Common SLAM interview questions, Part 1