Questions about code and dataset about ppdm HOT 12 CLOSED

1. We only evaluate our PPDM based on our rewriting evaluation script for V-COCO. Hourglass-104 achieved about 59.
2. Because we follow Objects as points to apply a data augmentation during training, the affine_transform function is used to generate the corresponding GT heatmap for the transformed input, e.g., scale crop aug, flip etc.

from ppdm.

The v-coco mAP is the same with the official one?

from ppdm.

We do not make sure it. We haven't conducted the evaluation in the official one. Its work is in the to-do list but not now, may next month. You can try it by yourself, and we have provided the corresponding training code for vcoco.

from ppdm.

Thanks, closing it

from ppdm.

1. We only evaluate our PPDM based on our rewriting evaluation script for V-COCO. Hourglass-104 achieved about 59.
2. Because we follow Objects as points to apply a data augmentation during training, the affine_transform function is used to generate the corresponding GT heatmap for the transformed input, e.g., scale crop aug, flip etc.

Could you share vcoco(Hourglass-104) training scripts or hyperparameters ? I tried but can not reach 59.

from ppdm.

All hyper-parameters are the same as hico-det's. Additionally, we only conduct an evaluation based on the non-official evolutional script, which cannot be compared to the methods evaluated by the official script.

from ppdm.

Thanks

from ppdm.

All hyper-parameters are the same as hico-det's. Additionally, we only conduct an evaluation based on the non-official evolutional script, which cannot be compared to the methods evaluated by the official script.

It is weird that I use the same hyper-parameters(lr3e-4 bs 31) in hico-det's training scripts to train vcoco, but still can not reach 59 using your vcoco_eval.py test script. I just get about 54.

from ppdm.

All hyper-parameters are the same as hico-det's. Additionally, we only conduct an evaluation based on the non-official evolutional script, which cannot be compared to the methods evaluated by the official script.

It is weird that I use the same hyper-parameters(lr3e-4 bs 31) in hico-det's training scripts to train vcoco, but still can not reach 59 using your vcoco_eval.py test script. I just get about 54.

I try to find the checkpoint, but I found all of them are deleted due to the space limitation. The 59 is got by the older version code, and you can try it based on this repo. Hope you can reproduce it.

from ppdm.

All hyper-parameters are the same as hico-det's. Additionally, we only conduct an evaluation based on the non-official evolutional script, which cannot be compared to the methods evaluated by the official script.

It is weird that I use the same hyper-parameters(lr3e-4 bs 31) in hico-det's training scripts to train vcoco, but still can not reach 59 using your vcoco_eval.py test script. I just get about 54.

I try to find the checkpoint, but I found all of them are deleted due to the space limitation. The 59 is got by the older version code, and you can try it based on this repo. Hope you can reproduce it.

Aha, I have just found the checkpoint, OneDrive or BaiduDrive? The only preserved one. However, I don't think the current training and test strategy is fit for the official vcoco evaluation method. It may be better to give a threshold to remove the non-interaction predictions or to take the non-interaction images into training.

from ppdm.

Thanks, both OneDrive and BaiduDrive is ok

from ppdm.

Thanks, both OneDrive and BaiduDrive is ok

OneDrive.

from ppdm.