search

megvii-research / video_analyst

A series of basic algorithms that are useful for video understanding, including Single Object Tracking (SOT), Video Object Segmentation (VOS) and so on.
MIT License
828 stars 176 forks source link

How to fine-tune SAT on a new dataset #199

Closed zzg-zzg closed 2 years ago

zzg-zzg commented 3 years ago

Thank you for your wonderful work, I tested the SAT on my own dataset but the results were not good enough. I guess it might be because my target category is not very common. I decided to fine-tune the pre-trained model on my dataset. What files and configurations do I need to modify? How to deal with the dataset(about the format)?

lzx1413 commented 3 years ago

The format of dataset is defined here https://github.com/MegviiDetection/video_analyst/blob/master/docs/TUTORIALS/VOS_SETUP.md, we did not format the dataset to the same schema, so you can find the dataset class here https://github.com/MegviiDetection/video_analyst/blob/master/videoanalyst/data/dataset/dataset_impl/davis.py and the config files lie in https://github.com/MegviiDetection/video_analyst/tree/master/experiments/sat/train

MARMOTatZJU commented 2 years ago

Closed as no further discussion. Feel free to reopen it again.