Question about the real time ability about the proposed approach

[KleistvonLiu]

Hi Martin,

thanks for your generosity. But I have a question about your algorithm. According to your paper, the measurements noise matrix at timestamps n+1 is estimated from the imu data at last N timestamps. But when I checked your code, I found that the Input of CNN is (2967,6) where 2967 is timestamps and 6 is acc and angular velocity. The output of CNN is the measurements covariance with a dim of (2967,2). In your implementation, you treated the first timestamp element (1,2) as the measurements cov at timestamp 1(assuming that we start from 0), I think it makes that the measurements noise matrix is estimated from the imu data at next N timestamps. At least I think if the input is data from n time points the output should not also contains estimates from n time points. If I misunderstand something, please tell me.

[Leooon726]

I have the same question. In the 2 convolution network of this code, the receptive field is 21(not sure), which means when we compute the covariance matrix at time t, the data from t-100ms to t+100ms is used(suppose the IMU update frequency is 100Hz), it seems unreasonable for online usage.