tfzhou / MATNet

Motion-Attentive Transition for Zero-Shot Video Object Segmentation (AAAI2020&TIP2021)

Home Page:https://www.researchgate.net/publication/343623463_MATNet_Motion-Attentive_Transition_Network_for_Zero-Shot_Video_Object_Segmentation

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Missing Files

MSiam opened this issue · comments

Hello thanks for your code

When I try to train MATNet after performing all the steps for dataset preparation. I still find the code is looking for these two json files can you please share them?
"train-train-meta.json"
"train-val-meta.json"

Thanks

Hi, these files should be provided by the Youtube-VOS dataset.

I didnt find them with YoutubeVOS 2018 dataset

I did find them however in RVOS repo they do provide https://github.com/imatge-upc/rvos
I am assuming you followed the setup from AGS which followed these I am guessing?

Cause they are the "train/val splits within the train" split. YoutubeVOS provides the train and val splits, without redividing train split again.

I am sorry that I cannot remember the details, but you can change the dataloader to fit the annotations. I did not follow the setting in AGS, but you can also refer to its setting.

thanks for your reply I will. Just small question in the paper you do mention "In total, we have 14K training sam-
ples, basically matching AGS (Wang et al. 2019b)" so it is safe to assume you are following AGS training setup I mean and what data you are using so I can refer later to that?

Yes. Both MATNet and AGS use about 14K data for training. However, the training datasets they used are different. For MATNet, youtubevos and davis are used for training.

I see thanks for your help