Do you know what caused this problem when I was running your code?
OSError: /GPUFS/.conda/envs/tor17/lib/python3.6/site-packages/torch_scatter/_version_cuda.so: undefined symbol: _ZN3c105ErrorC1ENS_14SourceLocationESs

Is all static images datasets for Pre-training the training set or both training and val set?
Is static images datasets COCO for Pre-training the coc2014 training set or coco2017 training set.

The offical YouTube-VOS dataset contains the valid set and the valid_all_frames, have you use the valid_all_frames or only the valid for the testing?

Hi,

Thank you for your nice work! I found the long time video dataset here.

Please tell me how to evaluate the results? Because the Annotation file is not complete.

Are there any Settings to reduce memory consumption?When I run with a 2080ti, I always report running out of memory.Still think your code doesn't support multiple GPUs?

How to calculate iou of test result on long video? Can you provide this test script?

Hi,

Ok so, I'm trying to train a custom dataset, either from scratch or from the checkpoint.
When I run:
python train.py --level 0 --dataset dataset_path\ --seed 1 --lr 1e-5 --scheduler-step 3 --new --total-epoch 12 --log --budget 1873920

OR:

python train.py --level 0 --dataset dataset_path\ --seed 1 --lr 1e-5 --scheduler-step 3 --resume checkpoint_path\checkpoint.pth --total-epoch 12 --log --budget 1873920

I get the ValueError mentioned in the title.
I'm guessing this is because it's looking for COCO, ECSSD, MSRA10k, PASCAL-S or PASCALVOC2021 and is only getting my custom dataset, but I would've thought it should still work given the file directories are set out as required.

My dataset_path folder has an Annotations folder, a ImageSets folder and a JPEGImages folder and then inside each of those are folders 0, 1 and 2 with the rest of the files contained in those just like the DAVIS dataset is set out.

What have I done wrong? Really looking forward to using this algorithm :)

In your paper, you refer to the resolution of backbone output from resnet.layer3, which output stride is 8. But in your code, you don't change original resnet50's stride, leading to a 16 of output stride.

coco 2014?

2 bugs in unify_pretrain_dataset.py

1, For MSRA10K
line 67-68, the directory name 'MSRA10K_Imgs_GT' is misspelled as 'MARA10K_Imgs_GT'

2, For COCO
line 167, 178, the directory name for the annotations in the downloaded file (2021.12.27) is 'annotations 2' instead of 'annotations'

Is this "dataset.txt " created by unify_pretrain_dataset.py or other ".py" file you didn't release?
PS: glob is built in package in python, no need to state in "requirements.txt".

Now, I'm fine. The "dataset.txt " contains dataset names like this

COCO
ECSSD
MSRA10K
PASCAL-S
PASCALVOC2012

Hi, thanks for your great work! What's the difference between p and q here?What do q and p stand for?

Is all static images datasets for Pre-training the training set or both training and val set?
Is static images datasets COCO for Pre-training the coc2014 training set or coco2017 training set.

I finished to collect five static image datasets, according to your readme.

11/08 18:02:44 114763 imgs are used for PreTrain. They are from ['COCO', 'ECSSD', 'MSRA10K', 'PASCAL-S', 'PASCALVOC2012'].
pretrain dataset length is  114763
11/08 18:02:44 Load level 0 dataset: 114763 training cases.
11/08 18:02:51 Random seed: 1604829771

pretrained dataset has 114763 images, almostly close to "136032" in your paper. But pretrain time is about 12 hours, which is largely far way from "days to week" in your README. May I make some mistakes?

Main Train:   1%|               | 879/114763 [05:07<10:48:59,  2.92it/s, loss=0.38932 (0.64092 0.56776)]

Hello, thanks for sharing such a complete code.

Maybe due to my inadequate understanding, I have some questions about the train process, I hope to get your answers.
In the dataloader DAVIS_Train_DS of training, each time, 6 randomly shuffled images are sampled from this video sequence, as mentioned in the paper [4.2 Main training on the benchmark datasets]. But in DAVIS_Test_DS, the images are read in order.

I want to know, what is the purpose of random sampling? Are there any obvious advantages to this approach?
If I read five or six consecutive images randomly, e.g. [random_id : random_id + 6], will there be any negative effects?

Thanks.