Adding a new category will affect the detection of other categories

[ease-zh]

I have tried your online demo, it's a great work.
But I found that adding a new category will affect the detection of other categories. For example, when I just detect "person", it works great,

but, when I add "light", it gets worse.

And when I use "traffic light" instead of "light", it works fine. And even I add a not existed "dog", it can still work.
I'm wondering, do categories affect each other? Or "light" is some magic word? Or "light" is just out of the seen vocabulary table?

[wanghao9610]

Hello @ease-zh , the input category names will affect the output results due to the language-aware selective fusion mechanism. Different category names for the same thing can also influence the output results. In your case, "light" has multiple meanings, while "traffic light" has a more specific meaning. I suggest trying different name choices in your application and selecting the one that works best.