Why do you use segmentation labels instead of bounding boxes in the COCO dataset?

[gongyan1]

Hello, This is indeed a fascinating piece of work, and I appreciate the outstanding contribution you have made to the community. However, I have a question. I noticed that the labels in this repository use segmentation instead of bounding boxes. Could you explain the reason behind this choice? If my dataset is labeled with bounding boxes, would it still be compatible? Moreover, were the performances of other methods compared in the article also trained with segmentation labels? To my knowledge, both v7 and v5 have used bounding box labels.

[LZHgrla]

Hi! @gongyan1 We use the same loss from yolov7, and do not involve seg labels.

Could you show the related code about the seg labels?

[gongyan1]

@LZHgrla Dear Lin,

Thank you for reaching out. However, when I downloaded the COCO labels mentioned in your readme, I found that they are not stored as four values of x, y, w, h as expected but rather as many values, and the number of values per line is inconsistent. Therefore, my question is, given that one txt file corresponds to one image and each line corresponds to one object, if these are bounding boxes, they should all be in the common format of class_id x y w h.

[LZHgrla]

@gongyan1 The COCO labels mentioned in our README are for downloading segmentation labels. However, during the training process, the segmentation labels will be processed into bounding box labels.

https://github.com/VDIGPKU/DynamicDet/blob/4dbb9b258e8052f9ab8668a37387e55a88fa6f4b/utils/datasets.py#L488-L496

You can also directly download bounding box labels here, https://github.com/ultralytics/yolov5/releases/download/v1.0/coco2017labels.zip