Train on Custom Images

[YubeiZheng]

Thank you very much for sharing, and I want to train my own images but don't know how to annotate relations. Could you please tell me what tools to use or how to mark the relationships between objects? Thank you very much.

[Jingkang50]

Dear Yubei,

Thank you for your question. The tool we use is simply a text annotation interface (supported by SenseTime Annotation Software, but it should be very simple that you can easily find alternatives elsewhere). We do some preprocessing on the image like here, i.e., put masks on the image with their tags (like person-0, umbrella-3). Then we ask the annotators to write down / correct relations, such as: 0 holding 3

Hope it helps.

[YubeiZheng]

Thank youfor your reply, it helps me a lot.