About the generalization capability of the proposed De-Bias Net

zhangkunyi / DIDO

Project page of 《DIDO:Deep Inertial Quadrotor Dynamical Odometry》

46 stars 6 forks source link

Open LarryDong opened 3 months ago

LarryDong commented 3 months ago

Hi, thanks for sharing the codes.

After reading your paper, I have some questions:

In your paper, you trained the network on Blackbird and VID dataset, and tested on "unseen simple trajectoris". Are these test trajectoris in Blackbird and VID dataset?
Do we need to first train the network on our own platform/sensor setup, if we've changed the platform or IMU mounting?

Thanks.

JIANG-CX commented 2 months ago

As stated in the paper, we did not train the network on the Blackbird or VID datasets; instead, we trained and tested it entirely on our own captured dataset, which is located in the dataset file.
Yes, the network should be trained on each specific platform/sensor setup.