xuelunshen/gim

train with my own video

fighterzzzh opened this issue · 4 comments

fighterzzzh commented

Hello,

I would like to ask if I need to train with my own video? My video contains my own scenes, but the camera might not be able to recognize some moving objects in the scene.

Thank you very much!

xuelunshen commented

Hello, I didn't fully understand the question you wanted to ask. Could you please be more specific and clear about it?

fighterzzzh commented

您好，我不太明白您问的问题，您能具体说清楚一点吗？
Sorry, I may not have explained it very clearly. Here is the situation: I am using SuperPoint+LightGlue for matching, but the performance on containers is not very good. Using the model you provided directly also did not yield good results. What I mean is, would it be better if I collect videos related to containers and use GIM for training? Are there any specific requirements for the videos? The videos I collect usually come from fixed-position cameras.

xuelunshen commented

Thank you for your explanation. Generally speaking, training with videos from your real application scenarios and then adapting to it will yield better performance, which is worth trying. Even for fixed-position cameras, GIM can be used to create training data. However, we haven't finished organizing all the code for GIM yet, so you'll need to wait a bit longer for that. Have you tried GIM_DKM? What I'm more concerned about is whether GIM_DKM also fails to your real application scenario?

fighterzzzh commented

Thank you for your explanation. Generally speaking, training with videos from your real application scenarios and then adapting to it will yield better performance, which is worth trying. Even for fixed-position cameras, GIM can be used to create training data. However, we haven't finished organizing all the code for GIM yet, so you'll need to wait a bit longer for that. Have you tried GIM_DKM? What I'm more concerned about is whether GIM_DKM also fails to your real application scenario?

Thank you for your response. Dense and semi-dense methods like GIM_DKM indeed produce better results and almost solve the problem. However, they consume too much memory and are too slow, which are the issues I am facing.