Shoufa Chen

Hi, Thanks for your interest in our work. The raw signals~(eg, image pixel, ground truth boxes coordinates) should be scaled to some range before combination with noise. Taking image generation,...

Hi, The set prediction loss contains one item $\mathcal{L}_{L1}$ defined [here](https://github.com/ShoufaChen/DiffusionDet/blob/main/diffusiondet/loss.py#L196), which measures the mean absolute error (L1 distance) between each element in the ground truth boxes and predicted boxes.

Hi, Thanks for your interest in our work. We've checked it and found similar results. The benefit of the diffusion model for object detection comes from two aspects: (1) random...

Hi, Thanks for your interest. It is a detectron2-related problem. How did you install detectron2?

Hi, Thanks for your interest in our work. Here is the DiffusionDet-COCO-Res50 training log: [log.txt](https://github.com/ShoufaChen/DiffusionDet/files/10109448/log.txt)

Hi, @bravezzzzzz This is an outdated log. Please wait for about 1 week. We will provide an updated log, which achieves 45.8 AP, ResNet50 with 300 evaluation boxes and a...

Hi, Thanks for your interest. We agree with you that cloned queries will produce the same predictions, and NMS is needed for post-processing. Anyway, in the best situation, DETR would...

Hello everyone, Thanks for your interest in our work. We have fixed the DETR baseline with the dynamic box setting by adopting NMS when $N_{eval} > N_{train}$. The updated results...

Hi, Thanks for your interest. Our implementation is based on Detectron2. So, I think you can follow https://github.com/facebookresearch/detectron2/blob/main/docs/tutorials/datasets.md#use-custom-datasets.

Hi, Thanks for your interest in our work. We didn't try to denoise the confidence in this version. We think it is an interesting idea to diffuse the confidence score.