kitti_odometry_eval

Brief

KITTI odometry evaluation contains CPP files for loading estimated and ground truth poses, and then computing the following metrics: average translation error [%] & average rotation error [deg/m].

Compile

The downloaded eval files can be compiled using c++ compiler (clang, c++, g++, mingw, etc.). I ran the following to compile on Linux:

g++ -O3 -DNDEBUG -o evaluate_odometry evaluate_odometry.cpp matrix.cpp

Note: To evaluate on the ground truth poses that we possess changed line 426 from: for (int32_t i=0; i<11; i++) { to:

# PA poses available
for (int32_t i=PA; i<PA+1; i++) {

# Example -> poses available "05.txt, 06.txt"
for (int32_t i=5; i<7; i++) {

Usage

To evaluate on the ground truth poses that possess. The compiled program (evaluate_odometry) can then be run:

/evaluate_odometry <result_sha>

# result_sha is the sub-directory containing the estimated poses
# poses must be in relative directory "results/<result_sha>/data"
# with each sequence as "XX.txt" inside the data folder

result_sha
├── data
│   └─── 00.txt
├── ...
└── ...

# Example
/evaluate_odometry 00

Full Structure

Evaluation creates the following files:

• errors: space-delimited file with rows
  • first frame number
  • rotation error
  • translation error
  • sequence length (path length) [m]
  • speed [km/h]
• plot_error: Plots of error for each sequence
  • path length vs. rotation error
  • speed vs. rotation error
  • path length vs. translation error
  • speed vs. translation error
• plot_path: x and z coordinates [m] for the path of the ground truth and odometry output. The sequence start is also marked. Files are included that display coordinates of all points for each sequence.
• stats.txt: average translation error followed by average rotation error (i.e. comparing gt to gt yields 0.0 and 0.0...).

results
├──00 // * example <result_sha> 
│    ├── data
│    │   └─── 00.txt
│    ├── errors                   
│    │   └─── 00.txt
│    ├── plot_error               
│    │   └─── <pre_sub>.<format>      
│    │         // 1. <pre_sub> 
│    │         // 1.1 pre [avg, <num>] 
│    │         //      1.1.1 (avg: average, <num>:estimated pose)
│    │         // 1.2 sub [rl, rs, tl, ts] 
│    │         //      1.2.1 (r:rotation, t:translation, s:speed, l:path length)
│    │         // 2. <format> [ eps, gp, pdf, png, txt]
│    │         // 3. * example: avg_rl.pdf ... 00_rl.pdf ...                        
│    ├── plot path                                      
│    │   └─── 00.eps
│    │   └─── 00.gp
│    │   └─── 00.pdf
│    │   └─── 00.png
│    │   └─── 00.txt
│    stats.txt  
├── ...
└── ...

Reference

• KITTI odometry development kit [Official devkit]
• ORB-SLAM-Experiments-and-KITTI-Evaluation