Hello, I'm trying to train the model with only dice loss (i.e no BCE loss) but the dice loss does not improve as you can see in the figure below. Pink is dice only, blue is dice+BCE. These are the resulting masks, where the dice-only-masks (left) are completely uniform and does not improve at all....

When executing “python interactive_demo.py --video [path to the video] --num_objects 15” commands，You can only enter 9 at most through the keyboard，Both 0 and 1 are entered and cannot be recognized as 10

I used the default parameters in the code, first used two 2080ti to train s0, then used two v100 to train s3, and tested the results of 100K-110k iterations. The best result of Davis 2017 val was 85.3% (at 110K), which is lower than what you wrote in the paper (86.2). Do you know what caused this re...

" main training at 107K iterations leads to the best result (full training is 110K). "(see https://github.com/hkchengrex/XMem/blob/main/docs/TRAINING.md) Which training process is this sentence refer to, s012 or s03? The iterations of main training is 150K in s2 and 100K in s3 in your code, neither...

Hello ! thanks for sharing your great job! I noticed your model also achieved a promising result in an out-of-domain case. Have you tested your model's performance in other domains, e.g. infrared images? And how about the robustness of your model in the out-of-domain case? Thank you !

I would like to ask about the equipment you use for training and the time-consuming of each stage. How many 2080ti or v100 did you use for training? How many days does each stage take?

Hi, I've been reading the code in detail, and found something that looks like unintended behaviour. In inference/inference_core.py#L65, we create the new hidden state in the decoder using HiddenUpdater module. Now, if we have our deep_update_sync disabled, then our is_normal_update is always True, t...

(XMem) [email protected]:~/CJQ/XMem-main/XMem-main$ CUDA_VISIBLE_DEVICES=2,3 python -m torch.distributed.run --master_port 25763 --nproc_per_node=2 train.py --exp_id retrain --stage 03 WARNING:main: Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system bein...

(Xmen) [email protected]:/work/user3/wxq/XMem-main$ CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.run --master_port 25763 --nproc_per_node=2 train .py --exp_id retrain --stage 03 WARNING:main: Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your sys...

Hi! I'm been doing some experiments and have the following use-case: We run the model in the regular way on a full video The user manipulates the memory in some way (doesn't matter how exacty) We want to re-run the predictions on all the frames (excluding those with annotated masks provided, obviou...