dynamic KITTI sequence mentioned in your paper seems to be inconsistent.

[WYKXD]

Hello, after visualizing the dynamic KITTI download link you gave, the qualitative result picture of the dynamic KITTI sequence mentioned in your paper seems to be inconsistent. Did you use other sequences when testing？

[SecureSheII]

Dynamic KITTI is a strict subset of KITTI so all images you find in Dynamic KITTI are copied from KITTI. The image quality should be the same as KITTI as well as images shown in the paper. May I ask what "qualitative result picture" are you referring to?

[WYKXD]

Hello, attached are the pictures in your paper and the evaluation results on Dynamic KITTI using the evaluation code published on GitHub. The GT lines in KITTI in your paper are inconsistent with the 01,03,04,10 I drew on Dynamic KITTI. Meanwhile, In the given Dynamic KITTI, the sequence like 01 has two folders, 0 and 1. Which folder is used for evaluation? 逐梦 ***@***.***  

…

------------------ 原始邮件 ------------------ 发件人: "castacks/DytanVO" ***@***.***>; 发送时间: 2024年1月23日(星期二) 上午10:43 ***@***.***>; ***@***.******@***.***>; 主题: Re: [castacks/DytanVO] dynamic KITTI sequence mentioned in your paper seems to be inconsistent. (Issue #18) Dynamic KITTI is a strict subset of KITTI so all images you find in Dynamic KITTI are copied from KITTI. The image quality should be the same as KITTI as well as images shown in the paper. May I ask what "qualitative result picture" are you referring to? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[WYKXD]

Are the KITTI evaluation results in your paper obtained from the KITTI dynamic sequences you gave? I used the code and weights you published to evaluate and got inconsistent results, especially on the sequence of 04_0, it is best to get only 0.85

[SecureSheII]

Have you used the correct model weights for KITTI? On 04_0 which has 36 frames in total, the ATE is 0.0180 using the finetuned model and 0.0206 if loading the flow model and pose model separately. Regarding sequence 01, we report the result of 01_1 in our paper, the one with 133 frames in total. For the other sequences, we use 00_1 (28 frames), 01_1 (133 frames), 02_1 (67 frames), 03_0 (31 frames), 04_0 (36 frames), 07_1 (136 frames), 08_0 (40 frames), 10_1 (59 frames).

If you still get inconsistent results, please provide instructions on how to reproduce the result you got.

[WYKXD]

Thank you very much

[SecureSheII]

Thanks to @Xie-HIT, the reason might be found, which is on our end. Please refer to #20 (comment)