Skip to content

REPORT.md

Latest commit

 

History

History
36 lines (25 loc) · 1.46 KB

REPORT.md

File metadata and controls

36 lines (25 loc) · 1.46 KB

SFND 2D Feature Tracking

Mid-Term Report

MP.1 Data Buffer

Remove the first element of the vector, if the length of the vector is greater than the desired length, see MidTermProject_Camera_Student.cpp line 97-99

MP.2 Keypoint Detection

Use the string "detectorName" MidTermProject_Camera_Student.cpp line 50 to select the detector

MP.3 Keypoint Removal

use the cv::Rect::contains(...) method to check if the point lies inside, see lines 142-152 in MidTermProject_Camera_Student.cpp

MP.4 Keypoint Descriptors

Use the string "descriptorName" MidTermProject_Camera_Student.cpp line 54 to select the detector

MP.5 / MP.6 Keypoint Matching

Use the strings matcherType (line 57) and selectorType (line 59) to select the different types. Also see inline comments.

MP.7 / MP.8 / MP.9

Spreadsheet: https://docs.google.com/spreadsheets/d/1TBnmddn5DbPttQTVZIm9n1Z5tD7-qPWA2pFfM1eqjDs/edit?usp=sharing

Conclusions

  • BRISK detector/descriptor is the slowest
  • SIFT, AKAZE detectors/descriptor are too slow to run at the rate of the camera
  • FAST (detector) + BRIEF (descriptor) and FAST+ORB are the fastest combinations
  • HARRIS detector is also quite fast, but provides only a limited number of keypoints
  • Regarding time, SHITOMASI+BRIEF/ORB, HARRIS+BRIEF/ORB, FAST+BRIEF/ORB, ORB+BRIEF/ORB are okay
  • SHITOMASI and FAST detect keypoints quite well distributed over the backside of the car

Recommendation:

  1. FAST + BRIEF
  2. FAST + ORB
  3. SHITOMASI + BRIEF or ORB