Averaging GPS segments

RESULTS
6.6.2019

RankMethodTrainingTestingDifferenceLengthPointsTime
A68.5%62.2%6.3%99%9882%30 min
1B67.1%62.0%5.1%99%89%seconds
2C70.4%61.8%8.6%101%83%seconds
3D68.0%61.8%6.2%99%83%seconds
E68.3%61.7%6.6%99%145%30 min
4F66.6%61.5%5.1%100%70%seconds
5G67.4%61.2%6.2%100%107%10 min
6H66.6%61.2%5.4%102%205%seconds
7I68.1%60.9%7.2%99%67%seconds
MEDOIDS
DTW57.3%55.3%2.0%97%159%~ 1 hour
ERP58.0%55.9%2.1%97%182%~ 1 hour
Euclidean58.3%54.8%3.5%97%175%~ 1 hour
Frechet59.9%55.2%4.7%97%210%~ 1 hour
Hausdorff61.2%56.4%4.8%99%121%~ 1 hour
IRD61.9%56.7%5.2%98%169%~ 1 hour
LCSS*59.0%54.4%4.6%99%202%~ 1 hour
EDR*58.5%54.5%4.0%99%210%~ 1 hour
C-SIM*58.3%54.8%3.5%98%190%~ 1 hour
HC-SIM*58.9%55.3%3.6%98%190%~ 1 hour
CellNet47.4%53.8%-6.4%64.8%144%seconds
CellNet°64.7%61.2%3.5%96.3%144%seconds
* threshold was set to 5% of the normalized space side length
° order is given
We evaluated using a novel similarity measure called HC-SIM (Hierarchical Cell Similarity). It is a parameter independent variant of C-SIM measure which uses a grid to assess similarity of trajectories. Source code is available here.

Submitted methods are evaluated at the top (A-I). They are followed by Medoids computed using different trajectory similarity measures act as a baseline. The averaging step described in the CellNet paper is also shown for reference.

These results are now documented in this paper:
Pasi Fränti and Radu Mariescu-Istodor (2020) "Averaging GPS segments competition 2019" Pattern Recognition.
Test set is now public and can be downloaded from here.