search

trqminh / aistron

Amodal Instance Segmentation Toolbox and Benchmark
Apache License 2.0
22 stars 1 forks source link

Questions about custom dataset #3

Open quanw8781 opened 1 year ago

quanw8781 commented 1 year ago

Thank you for your excellent work and I have some questions. Is it possible to train and predict using my own data, and if so, what should I do about it? I'm using the Labelme tool to label the images. How can I convert the json files to COCOA format or any other format that works? Looking forward to your reply.

trqminh commented 1 year ago

Yes, it's possible to train with your custom data. However, we're currently not supporting training with custom dataset directly. You can still work around by doing the following steps:

I'll try to add custom dataset support in the near future, but in the mean time, you can work around based on the above steps.

binh0804 commented 5 months ago

Thank you for dedicating the time to this, hope to see your reply soon. Sorry, can you share how labelme (or another label tool) can label occluded part of objects? I am preparing custom dataset based on KINS dataset annotation format and face to some problem. There are some fields that I don't know how to label such as "a_segm","i_segm" in update_test_2020.json or "amodal_full" in instances_train.json. Or, I have to edit the annotations file manually?

CHN-001 commented 3 months ago

感谢您花时间做这件事，希望很快收到您的回复。 抱歉，您能分享一下 labelme（或其他标签工具）如何标记物体的遮挡部分吗？我正在根据 KINS 数据集注释格式准备自定义数据集，但遇到了一些问题。有些字段我不知道如何标记，例如 update_test_2020.json 中的“ a_segm ”、“ i_segm ”或 instance_train.json 中的“ amodal_full ”。或者，我必须手动编辑注释文件？

你好，我遇到了同样的问题，我现在有COCO格式的自建数据集，或者是labelme标注生成的json格式文件，但是这些文件没有关于Amodal的信息，这导致了在复现模型的时候无法评估非模态分割的性能。我应该如何制作COCOA格式呢？或者能否将已有的COCO格式的数据转换成COCOA格式的数据？

binh0804 commented 3 months ago

(Sorry for reply in Vietnamese because you use Chinese) Nhóm chúng tôi thực hiện các bước tương tự theo hướng dẫn sau: https://git.wur.nl/blok012/sizecnn/-/blob/master/ANNOTATE.md?ref_type=heads và sau đó chạy file labelme_to_orcnn.py (của repository đó). File annotation dưới dạng json lúc này đã có thể dùng với prepare_cocoa.py của aistron. Chúc may mắn!

CHN-001 commented 3 months ago

(Sorry for reply in Vietnamese because you use Chinese) Nhóm chúng tôi thực hiện các bước tương tự theo hướng dẫn sau: https://git.wur.nl/blok012/sizecnn/-/blob/master/ANNOTATE.md?ref_type=heads và sau đó chạy file labelme_to_orcnn.py (của repository đó). File annotation dưới dạng json lúc này đã có thể dùng với prepare_cocoa.py của aistron. Chúc may mắn!

Thank you very much indeed. You have been a great help to me. I will try to follow your method, but before I try, I have a few questions to ask you. First, if half of my tagged object is obscured and half is visible, should I label the visible half as visible and the whole as amodal? Second, if the entire amodal object is labeled with polygon, will human imagination affect the performance and effect? I will try it first. If I encounter any problems, I hope you can help me again. I really appreciate your help!

binh0804 commented 3 months ago

First: In my case, crystal, I label the visible part with class "crystal_visible" and set tag 1, and I imagine the complete shape of this crystal => label it "crystal_amodal" and set tag 1 also. Repeat to crystal 2,3,4,5, etc. Second: yes human imagination affect the performance. Of course, the machine learned what you trained it. You can contact me at binh08042002@gmail.com, I am also doing graduate project ơn this topic.

CHN-001 commented 3 months ago

First: In my case, crystal, I label the visible part with class "crystal_visible" and set tag 1, and I imagine the complete shape of this crystal => label it "crystal_amodal" and set tag 1 also. Repeat to crystal 2,3,4,5, etc. Second: yes human imagination affect the performance. Of course, the machine learned what you trained it. You can contact me at binh08042002@gmail.com, I am also doing graduate project ơn this topic.

Thank you very much for your answer, which is very clear and has solved my problems. I will contact you by email if I have any questions in the future! Thank you for your help!