luokai0 Computer Vision Robotics Expert
Before Starting
- RGB, depth, or stereo camera?
- Detection, pose estimation, or SLAM?
- Classical or deep learning approach?
Core Expertise Areas
Camera Models
Pinhole camera: projects 3D point to 2D image plane through focal point. Intrinsic matrix K: focal lengths and principal point, 5 parameters. Distortion: radial and tangential lens distortion, correct before processing. Stereo camera: two calibrated cameras, disparity gives depth. Depth cameras: structured light or time-of-flight, direct depth measurement.
Feature Detection
Corners: Harris detector, stable under rotation, used for tracking. SIFT: scale and rotation invariant features, robust but slow. ORB: binary features, fast, used in real-time SLAM. Feature matching: descriptor distance, ratio test, RANSAC for outlier rejection.
Object Detection
Classical: HOG plus SVM, sliding window, deformable parts model. YOLO: single-shot detector, real-time, divides image into grid. Faster R-CNN: region proposal network, accurate but slower. Segmentation: semantic labels per pixel, instance segmentation separates objects.
Visual SLAM
SLAM: simultaneous localization and mapping, build map while tracking position. Visual odometry: estimate motion from consecutive image frames. Loop closure: detect revisited places, correct accumulated drift. ORB-SLAM3: feature-based, monocular and RGB-D, open-source reference system. NERF and 3DGS: neural radiance fields for photorealistic scene reconstruction.
Best Practices
- Calibrate cameras carefully before any metric reconstruction task
- Use data augmentation to improve detection robustness
- Test perception system under lighting variation and occlusion
- Validate pose estimation with ground truth before deployment
Common Pitfalls
| Pitfall | Fix |
|---|---|
| Skipping camera calibration | Always calibrate, reprojection error should be below 1 pixel |
| Ignoring lighting variation | Test and train under diverse lighting conditions |
| Assuming static scene in SLAM | Handle dynamic objects explicitly |
| Overconfident pose estimates | Always provide uncertainty with pose estimates |
Related Skills
- path-planning-expert
- robot-kinematics-expert
- deep-learning-expert