How to fine tune a model to a custom dataset

[minocrafft]

Hi @drinkingcoder,

First of all, thank you for your wonderful research. I am interested in training the FlowFormer model in custom data. Currently, it seems that training codes are only provided for benchmark datasets such as chairs, things, sintel, kiti, etc. I have some video data, and I'm going to cut the video into frames and use it for training.

I'm just starting research in the field of optical flow, so I don't know much about it. Is there a code that processes custom dataset for training? Or do I have to make a custom dataset myself? An additional, How do I create .flo file?

Thank you in advance for your response!

P.S.) What does autoflow mean in code that handles the args.stage of the train code?

[drinkingcoder]

You could customize the data loader from chairs, or things. But I think the biggest challenge is to obtain the optical flow GT of the your videos.

[shruti18196]

You could customize the data loader from chairs, or things. But I think the biggest challenge is to obtain the optical flow GT of the your videos.

Can we use FlowFormer pre-trained models to generate GT for our custom dataset?

[SacherTorte280]

You could customize the data loader from chairs, or things. But I think the biggest challenge is to obtain the optical flow GT of the your videos.

Can we use FlowFormer pre-trained models to generate GT for our custom dataset?

Same need, I wish if you can have any progress.