Can the prompt for the first frame of the image used in running the model be a GT box instead of a mask?

[atoneyzoek]

Hello, your work is excellent! But I encountered a problem when trying your code. When running run_on_box_dataset.py, it reports an error: 'Error: Init mask on this path does not exist.' Does this mean that the first frame in the code must use a mask as a prompt? Is it possible to simply use my annotated GT boxes? If possible, please help me understand how to operate it so that I can run your model using only the GT box. Thank you!

[alanlukezic]

Hi, we are glad to hear that you like our work. The script you are trying to run (run_on_box_dataset.py) assumes that you have already computed initialization segmentation masks for all sequences in advanced. We have already done this and provide the init masks to assure reproducible results. To know more about this, please see here
If you don't have the initialization masks, you can run the run_bbox_example.py which asks the user to draw the init bounding box and then runs the tracker - see here for more info. I hope that this answers your question.

[atoneyzoek]

Thank you for your reply! I know that run_bbox_example.py requires manually drawing an initial bounding box before it can run. What I mainly want to know is whether your model can run with only the GT boxes? Because I have annotated some sequences myself (their GT boxes are all labeled, but the initial masks are not), I want to know if it's possible to use your code without annotating the initial masks？ I think labeling a GT box is more in line with everyday usage and relatively simpler. Thank you!

Hi, we are glad to hear that you like our work. The script you are trying to run (run_on_box_dataset.py) assumes that you have already computed initialization segmentation masks for all sequences in advanced. We have already done this and provide the init masks to assure reproducible results. To know more about this, please see here你好，很高兴听到你喜欢我们的作品。你试图运行的脚本（run_on_box_dataset.py）假设你已经提前计算好所有序列的初始化分割遮罩。我们已经做过这些，并提供了初始掩码以确保结果的可复现性。想了解更多，请参见这里 If you don't have the initialization masks, you can run the run_bbox_example.py which asks the user to draw the init bounding box and then runs the tracker - see here for more info. I hope that this answers your question.如果你没有初始化遮罩，可以运行一个 run_bbox_example.py，要求用户绘制初始化边界框，然后运行追踪器——更多信息请见这里 。希望这能回答你的问题。

[alanlukezic]

I think the easiest way would be to modify the run_bbox_example.py script. In particular: change the lines 25-26 in this script by replacing the BoxSelector with the code that reads the groundtruth init file (first frame bounding box) saved on the disk.