search

DragonEmperorG / QDeepOdo

Reproduce the paper “A Novel Deep Odometry Network for VehiclePositioning Based on Smartphone” method
2 stars 1 forks source link

【讨论】调整网络输出结构从多对多到多对一，在经过尽量小的窗口大小后能够预测出较准确的前向速度 #3

Open DragonEmperorG opened 10 months ago

DragonEmperorG commented 10 months ago

训练配置

train validation window
35 4 60 s

loss为切片序列的MSE

训练过程

image

DragonEmperorG commented 10 months ago

训练配置

train validation window
35 4 60 s

CNN+RNN网络 loss为切片序列最后最后一个历元的MSE

训练过程

image

训练结果

2020-08-13-21-42-us-ca-mtv-sf-280_pixel5_forward_velocity_result 2021-04-02-20-43-us-ca-mtv-f_pixel5_forward_velocity_result 2021-08-24-20-32-us-ca-mtv-h_pixel5_forward_velocity_result 2021-12-09-17-06-us-ca-lax-e_pixel5_forward_velocity_result

讨论

基于深度学习方法直接估计前向速度泛化性存在问题，

DragonEmperorG commented 10 months ago

训练配置

train validation window
31 8 1 s

CNN网络 loss为切片序列最后最后一个历元的MSE

训练过程

image

训练结果

2020-08-13-21-42-us-ca-mtv-sf-280_pixel5_forward_velocity_result 2021-04-02-20-43-us-ca-mtv-f_pixel5_forward_velocity_result 2021-08-24-20-32-us-ca-mtv-h_pixel5_forward_velocity_result 2021-12-09-17-06-us-ca-lax-e_pixel5_forward_velocity_result