How to improve semantic differences between different objects

[jkff00]

The similarity matrix obtained based on a single query shows little semantic difference between the different objects in the scene.

[SunzeY]

yah, This is great problem. During our training, we find there is a trade of between "better local focus to distinct each object when test on object detection dataset like coco" and the "general recogonition ability like image classification accuracy on imagenet"... How to solve this is an open and may be a good research problem...

[SunzeY]

I believe this is depends on what you want to achieve, if you use alpha-clip for some relatively low related promblems instead of directly modeling object level semantics, it may work better than image level CLIP. This paper: IODA: Instance-Guided One-shot Domain Adaptation for Super-Resolution may help you find your way on this...

[jkff00]

Is this a strategy based on mapping the image dimension to a high-dimensional semantic space? If the mapping is accurate, it may only include the subject and its context within the mask, thereby increasing the direct difference between the objects.

…

---Original--- From: "Zeyi ***@***.***> Date: Sat, Jan 11, 2025 16:28 PM To: ***@***.***>; Cc: ***@***.******@***.***>; Subject: Re: [SunzeY/AlphaCLIP] How to improve semantic differences betweendifferent objects (Issue #67) I believe this is depends on what you want to achieve, if you use alpha-clip for some relatively low related promblems instead of directly modeling object level semantics, it may work better than image level CLIP. This paper: IODA: Instance-Guided One-shot Domain Adaptation for Super-Resolution may help you find your way on this... — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>