The VO results are bad when using newer pytorch

[eric-yyjau]

The KITTI VO results are bad when using the following pytorch version.

torch==1.5.1
torchvision==0.6.1

Follow the same scripts on KITTI sequence 10, I got the following results.

Sequence: 10
Translational error (%):  12.00700933424375
Rotational error (deg/100m):  5.758592589262757

The results are good under pytorch 1.2.
Why is it the case? Thanks.

[thuzhaowang]

We have yet not tested with pytorch 1.5. If you have any further findings about this version issue, please report here. Thanks.

[chaoxingxi]

The KITTI VO results are bad when using the following pytorch version.

torch==1.5.1
torchvision==0.6.1

Follow the same scripts on KITTI sequence 10, I got the following results.

Sequence: 10
Translational error (%):  12.00700933424375
Rotational error (deg/100m):  5.758592589262757

The results are good under pytorch 1.2.
Why is it the case? Thanks.

The figure is the result of use
python infer_vo.py --config_file ./config/odo.yaml --gpu [gpu_id] --traj_save_dir_txt [where/to/save/the/prediction/file] --sequences_root_dir [the/root/dir/of/your/image/sequences] --sequence [the sequence id] ----pretrained_model [path/to/your/model] python ./core/evaluation/eval_odom.py --gt_txt [path/to/your/groundtruth/poses/txt] --result_txt [path/to/your/prediction/txt] --seq [sequence id to evaluate]
why my result use the provide pretrained models look more bad?
@B1ueber2y @thuzhaowang

[B1ueber2y]

Have you switched to the required pytorch version? It is reported that the grid sampling layer might be inconsistent with newer pytorch (e.g. pytorch==1.5)

[chaoxingxi]

Thank you for your reply. I have use the required pytorch version

Is something wrong when I run the infer?

infer_vo.py --config_file ./config/odo.yaml --gpu 0 --traj_save_dir_txt ./results/prediction/pose_10.txt --sequences_root_dir /media/xxxxxx/disk4t/dataset/KITTI/odometry/dataset/sequences --sequence 10 --pretrained_model models/pretrained/kitti_odo.pth

then i plot the result like this

Another question
How convert the results from CC(or SfMlearner) to Translational error (%), Rotational error (deg/100m)
When i do this, i find the later part of the sequence is bad than preceding part, like this (CC in seq10)

A little different from the visualization in your paper

Thank you

[henry123-boy]

I also met this problem, we follow the same evaluation setting provided and test the pre-trained model in KITTI sequence 09, but the result is very bad?