search

CathIAS / TLIO

Tight Learned Inertial Odometry
Other
277 stars 69 forks source link

The vio-displacement is used as the regressed target, why can the velocity be ignored? #26

Closed DHNicoles closed 2 years ago

DHNicoles commented 2 years ago

Hi, If imu moves at a constant speed at different speeds, the imu_raw data will be the same. Then the neural network will output the same displacement while inferencing. So why the velocity could be ignored when training the network? I'm not sure if the question is clearly stated, thank you in advance.

CathIAS commented 2 years ago

Hi @DHNicoles, Thank you for the question and my sincere apologies for the delay in my reply. Your observation is correct. The removal of initial velocity from the input is to force the network to not learn the integration function which mathematically is known to us, but learn the human movement patterns instead. In pedestrian applications, these perfect constant speed situations do not naturally occur and what the network is looking for are features like steps and the way the steps are taken.

DHNicoles commented 2 years ago

Got it.