only use object detection? custom dataset do not have segmentation labels

[zhenshen-mla]

Thanks for the work. If only detection is used instead of both seg and det , will the mAP decrease? Can you estimate the magnitude of the decline?

[ZhangGongjie]

Thanks for your interest in our work! :)

Please note that we only perform object detection, i.e., only bounding box annotations are used during training.

[ZhangGongjie]

The keypoints in our work are automatically learned without explicit supervision.

[zhenshen-mla]

thanks Here's the code from the main.py

* Segmentation

parser.add_argument('--masks', action='store_true', help="Train segmentation head if the flag is provided") So how should the parameters be set? the practical task is to detect hidden dangers in the transmission scenario, and it is not clear whether it is better to turn on or off the mask. Note: Trial and error costs are very very high.

[ZhangGongjie]

Thanks.

This segmentation flag is only here due to historical reasons. We modify the code from the original DETR, and DETR can be extended for panoptic segmentation by setting this flag. However, in our SAM-DETR, this flag should never be set.