vita-epfl / deposit Goto Github PK

Hi authors,
Great work! The study mentions that post-processing can be used to enhance the human pose of any prediction model. Can the authors comment if it can be used for repairing key points. For example, if we have only 14 predicted key points in COCO format (from any model like MediaPipe/ YOLOv8), and 3 missing key points. Can the approach be used to predict the remaining key points? Thanks!

Hey authors,
I am very interested in your work. Could you please share your pre-trained checkpoints? I want to evaluate your model without training. In addition, your paper mentions that post-processing can enhance any posture prediction model. I wonder if it can be used in stochastic models. It seems that only determinative models are enhanced in this paper.

Thank you for your grest work!
I saw that you used the cosine noise scheduler in the paper, could you tell me what its values are?

Hey,

The approach mentioned in this paper seems very intriguing and useful. However, I wanted to test out the usage of just the pre-processing phase where the key points are repaired. Are there any particular steps that I can be pointed towards so that I am able to just observe the broken and repaired key points after observation is passed through the Pre-Process Diffusion block?

Hey, though the model training time is high, due to the computational complexity of denoising diffusion models, can the authors comment on the inference time of their approach? Does the model have high time complexity at inference or can it be implemented for real time inference?

I appreciate the effort and the code-sharing very lot.
The training process takes a long time when using an Nvidia RTX 3090 to train the model (both small and large).
Each epoch can take up to 2 hours for the large model when the batch size is set to 16 (because of memory constraints).
Are you having the same problems or using more powerful setups?
Are there any ways, as far as you're aware, to reduce the training period?
Thank you for taking the time to respond.

Hello, we are very interested in your work, and we would like to replicate your results.
On the Human3.6M dataset, we trained and evaluated the resulting model, achieving a prediction accuracy similar to that of the paper. However, on the AMASS dataset, the model we trained cannot reach the accuracy of the paper, and the gap is large. Our training parameters follow your code exactly. The final evaluation results are as follows:
Loss of training on AMASS datasets:

FDE on AMASS:

FDE on 3DPW:

We suspect it may be due to hyperparameter Settings or accidental problems. Can you provide training hyperparameters and chackpoint on AMASS and 3DPW datasets? It would be better if some suggestions could be made for our restoration work.

Hello

Thank you for providing your codes. But the dataset link seems broken. Is it possible for you to provide the dataset link?

Thank You