search

JonathonLuiten / PReMVOS

Proposal-generation, Refinement and Merging for Video Object Segmentation
MIT License
129 stars 25 forks source link

Fine-tuning the proposal_net for a new video #13

Closed hejm37 closed 4 years ago

hejm37 commented 4 years ago

Hi, I want to fine-tune the method on a new video and I've learnt from your ACCV paper that there are three parts that were fine-tuned: Category Agnostic Mask R-CNN (proposal_net), Refinement Network and ReID Network.

However, I did not find any code in the proposal_net folder that performs the fine-tune operation and I was wondering if you could provide any help? Thanks!

JonathonLuiten commented 4 years ago

All the code for performing fine tuning the proposal net is here in the repo. Specifically finetuning can be called using the right parameters of train.py. However it is a little hard to figure out exactly what parameters are needed.

Looking at some old code, I think this should work: train.py --logdir SOMETHING --agnostic --second_head --resume --davis --forward_dataset DAVIS --load SOMETHING--davis_name SOMETHING with the SOMETHINGs being replaced with what you want to do.

Also here is a very old, very badly coded script I used at some point in early 2018. No guarantee that this works out of the box, but should point you in the right direction.

mask.zip

hejm37 commented 4 years ago

Thanks for your pointer!