The given checkpoint do not match all the model, and it's hard to reproduce the result

[64327069]

The given checkpoint miss the parameters for the affine_mlp model in the nvp_simplified.py

[64327069]

the checkpoint is downloaded here:This folder contains a script for evaluation (eval_tapvid_davis.py) and the pre-trained weights which you can use to reproduce the exact result in the paper.

We found the rgb picture shape is 224x224, however，the original picture in davis is 480*800+ We can not reproduce the result in the original picture, neither the scaled picture.

could you please give the checkpoint with the affine_mlp module parameters for continue training

[64327069]

And could you provide one sequence of video in davis (for example: blackswan) and detailed training config? The sequence can be directly trained and reproduce the result. Because we do not know how to produce the 224x224 pictures and extract their flow features Best wishes

[qianqianwang68]

Hi, the checkpoints we released were trained without the affine_mlp, i.e., (use_affine is False), and there are other parameter differences as indicated in this file: args.min_depth = -0.5 args.max_depth = 0.5. For more details, please refer to this thread.

These checkpoints are for the TAP-Vid benchmark evaluation which contains 256x256 resolution videos, and that's why the resolution is 256x256. We don't have the checkpoints for all 480p DAVIS videos.

[64327069]

谢谢回复，想问一下抽取光流特征的时候也是先把原图全部缩放成256x256然后再调用preprocessing对么

[qianqianwang68]

对的 https://drive.google.com/file/d/1bgakkkfw_HLzAQsCVVm4Gse5Unt9fyHA/view?usp=drive_link 这里面checkpoints里面color folder的图片是256x256的可以直接用

[pubyLu]

作者您好！ 我没有找到整个模型完整的训练checkpoint，请问可以贴一个链接吗？我想用OminiMotion测试一些自己拍摄的数据集