search

hkchengrex / XMem

[ECCV 2022] XMem: Long-Term Video Object Segmentation with an Atkinson-Shiffrin Memory Model
https://hkchengrex.com/XMem/
MIT License
1.72k stars 191 forks source link

The dataset used in stage03 #109

Closed zht8506 closed 1 year ago

zht8506 commented 1 year ago

I found that validating the davis2017 results requires the use of youtube data, and I'm wondering which data exactly? I used youtube2018-train.zip but only get 85.5 JF (vs 86.2 in your paper) on davis2017？ Do I need to use youtube2018-train_all_frames.zip or youtube2019-train.zip? Thank you very much.

1334233852 commented 1 year ago

I tested my own dataset, why didn't there be JF printing，thank you

hkchengrex commented 1 year ago

We used DAVIS 2017 train + YouTubeVOS 2019 train following standard practices.

zht8506 commented 1 year ago

I tested my own dataset, why didn't there be JF printing，thank you

It needs to be validated separately to get the metrics, and no validation will be done during training.

zht8506 commented 1 year ago

Thank you for your reply, it addresses my question very well.

1334233852 commented 1 year ago

It needs to be validated separately to get the metrics, and no validation will be done during training.

Do you mean to write a separate Python file for validation metrics, which is not included in the source code? Sorry, I am a novice and asked these questions.

hkchengrex commented 1 year ago

@1334233852 You can use the tools listed here: https://github.com/hkchengrex/XMem/blob/main/docs/INFERENCE.md#getting-quantitative-results