Thanks for your wonderful work! Did you use the setting of stage 2 only when reporting the DAVIS+YouTubeVOS only(training data) score on YouTubeVOS2018 validation set? I retrained stage 2 but it only gains 82.6 on YouTubeVOS2018 validation set with no extra change.

Hello! First of all – thank you for open sourcing your work, and congrats on being accepted to ECCV 2022! Would you be interested in porting the interactive demo over to Hugging Face Spaces? We have documentation for creating Spaces, but I'm also more than happy to help you out if it's something you...

Hi~ I'm using this repo to test my custom video, but get CUDA out of memory error. The resolution of my video is 4k, video length is 60s at 25FPS. Which settings could I change to make it possible to run this excellent work? Many thanks~~

I would like to be able to pause when doing segmentation for a very long video. Is it safe to save MemoryManager (https://github.com/hkchengrex/XMem/blob/main/inference/memory_manager.py) to disk as a .pkl on disk, and reload it later to continue on the next frame? Does it contain everything that is...

Hi, I would like to ask if there is any code for calculating indicators during test, for example: J&F?

Hi @hkchengrex！I trained on two 2060s, the batch size was set as 2, and only changed the size of batch_size. It was found that the loss curve decreased first, reached the lowest point of 1 at about 10K, and then kept rising until it reached 3 at 100K. Is this a normal phenomenon? My loss curve is a...

Hi, could you please share the training loss curve? I want to see the decline rates and convergence values of dice_loss and ce_loss. Thanks!

It looks just randomly sampled from historical reference frames. Is it not fully open source yet? Or does this simple strategy work well? Anyway, thanks for your work.

Hi, I couldn't find it under inference, so I was wondering, is there a way to use a pretrained model and run inference on a video I have and see how the model performs? Thanks!

Hi @hkchengrex Since my gpu memory is limited, I can only train youtube and davis with batch size=4, do you think this will affect the accuracy?