cannot reproduce VO accuracy when I try to train the pose model on KITTI Odometry

[Kewenjing1020]

Hi, I've tried to train the pose model on KITTI Odometry by running sh scripts/train_resnet50_pose_256.sh , here's the accuracy I get:

Sequence: 9
Translational error (%):  11.425453298024262
Rotational error (deg/100m):  3.5656289967743824
ATE (m):  68.78972540413156
RPE (m):  0.1264095786586114
RPE (deg):  0.09296849856063301
Sequence: 10
Translational error (%):  9.954224810755791
Rotational error (deg/100m):  3.078602353206078
ATE (m):  16.70645792760232
RPE (m):  0.07903440253451217
RPE (deg):  0.09633480322105922

I didn't change anything in the code/script. What might cause this accuracy gap?

[JiawangBian]

Odometry is an accumulated sequentially from pairwise predictions, so the results using different models would be different even though these models show very similar accuracy on pairwise prediction.

[essamgoda]

Hi, I've tried to train my own model on KITTI Odometry for only 5 epoch, here's the accuracy I get:

Sequence: 9
Translational error (%): 93.34175477374941
Rotational error (deg/100m): 22.65699349729781
ATE (m): 193.29374882824112
RPE (m): 1.2917560700283468
RPE (deg): 0.6969629843516546
Sequence: 10
Translational error (%): 136.31391520845585
Rotational error (deg/100m): 21.387362426113125
ATE (m): 78.570022262845
RPE (m): 1.2763249421405172
RPE (deg): 0.5787483638322359

is the number of epochs that might cause this accuracy gap (also PoseResNet results are very similar for the same epochs)?