search

Atten4Vis / MS-DETR

[CVPR 2024] The official implementation for "MS-DETR: Efficient DETR Training with Mixed Supervision"
Apache License 2.0
81 stars 5 forks source link

We are very interested in your work. Could you please answer the following questions: #14

Open 2308700388 opened 2 months ago

2308700388 commented 2 months ago

What are the effects of not using a two-stage approach for one-to-many supervision? Is the one-to-many supervision prediction head used during model inference, or is it used for both training and inference?

ZhaoChuyang commented 2 months ago

Thank you for your interest. Regarding your questions:

What are the effects of not using a two-stage approach for one-to-many supervision?

We haven't tested one-to-many supervision with a one-stage approach. You can explore this by disabling the --two-stage flag in the script, which will switch to one-stage training.

Is the one-to-many supervision prediction head used during model inference, or is it used for both training and inference?

The one-to-many prediction head is only used during training and is not used in the inference.

Please feel free to reach out if you have any further questions.

2308700388 commented 2 months ago

thanks,