Shoufa Chen comments

Results 121 comments of


                                            Shoufa Chen

Regarding demo collab

Hi, Thanks for your interest. Here is an unofficial implementation https://colab.research.google.com/drive/1rA5SXuTx2pI6o7tWA6Ad5QRZn4a1ajMh#scrollTo=Sn5gWF3fhpf- Thank @345ishaan for sharing this colab.

Regarding demo collab

@chenqq1 hello，链接是完整的，可能需要开一下代理

how to train diffusiondet on my own datasets？

Please see https://github.com/ShoufaChen/DiffusionDet/issues/20.

checkpoint

Hi, You can set `SOLVER.CHECKPOINT_PERIOD 5000` to your desired value. FYI: https://github.com/facebookresearch/detectron2/blob/main/detectron2/config/defaults.py#L554.

Question about `fc_feature = fc_feature * (scale + 1) + shift` in head.py.

Hi, This is used for time embedding. We directly follow the original Diffusion model implementation: https://github.com/lucidrains/denoising-diffusion-pytorch/blob/main/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py#L179-L180.

[A Bug?] The diffused boxes as input may have negative coordinates.

Hi, Thanks for your interest. Yes, the RCNN head can accept negative coordinates.

show your command line

Hi, Thanks for your interest. You could find detailed instructions here: https://github.com/ShoufaChen/DiffusionDet/blob/main/GETTING_STARTED.md.

`clip_x_start` is not used

Good catch! Thank you. We implement the clip in https://github.com/ShoufaChen/DiffusionDet/blob/main/diffusiondet/detector.py#L171.

Have you tried Hungarian matching instead of optimal transport assignment?

Hi, Thanks for your interest. We tried Hungarian matching in our pilot experiments and found that assigning multiple predictions to ground truth is better for DiffusionDet.

Have you tried Hungarian matching instead of optimal transport assignment?

Hi, @dk-liang , Apologies for the delayed response. At the initial stage of the project, we conducted a comparison between OT and Hungarian, but we don't have any quantifiable data...