Closed yanghhx closed 2 years ago
The data for skeleton-dml is constructed like described in the paper (Figure 3 and 4). The simplest way to construct the data like this would be using the numpy reshape function with (-1,-1,3) as arguments.
Umm, I am sorry that I am still confused how the skeleton is related to the image representation. Could you please offer the related code?
At 2021-11-09 23:09:52, "Raphael Memmesheimer" @.***> wrote:
The data for skeleton-dml is constructed like described in the paper (Figure 3 and 4). The simplest way to construct the data like this would be using the numpy reshape function with (-1,-1,3) as arguments.
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.
I have another question. When I set the bachsize to 16, the last iteration comes to error and remind me that the dimension is wrong.
At 2021-11-09 23:09:52, "Raphael Memmesheimer" @.***> wrote:
The data for skeleton-dml is constructed like described in the paper (Figure 3 and 4). The simplest way to construct the data like this would be using the numpy reshape function with (-1,-1,3) as arguments.
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.
Dear author: I am sorry to trouble you but I'd like to know why your dataset is not "one-shot" because it is the whole NTU 120 datasets. However, your work is "one shot action recognition", so can you tell me the reason?
At 2021-11-09 23:09:52, "Raphael Memmesheimer" @.***> wrote:
The data for skeleton-dml is constructed like described in the paper (Figure 3 and 4). The simplest way to construct the data like this would be using the numpy reshape function with (-1,-1,3) as arguments.
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.
Umm, I am sorry that I am still confused how the skeleton is related to the image representation. Could you please offer the related code?
I'll try to give a short example on the representation generation soon.
I have another question. When I set the bachsize to 16, the last iteration comes to error and remind me that the dimension is wrong.
Not sure what is happening there. Usually during training, it should randomly sample the amount of batch_size samples and because of the shuffling I'm quite sure that it should always fill up to the amount of samples.
I am sorry to trouble you but I'd like to know why your dataset is not "one-shot" because it is the whole NTU 120 datasets. However, your work is "one shot action recognition", so can you tell me the reason?
I suggest you to reread the paper and the NTU 120 paper carefully (The one-shot protocol is presented in a later section "One-Shot Recognition on NTU RGB+D 120"). In the paper we follow the one-shot action recognition protocol which is also described in the NTU 120 datset. Loosely spoken, you train a model on a set of up to 100 known classes and use 20 classes with one reference sample from each that have not seen during the training process for experiments. Initially, this might be a bit more tricky to understand, as the problem is not framed in a supervised classification setting.
To elaborate on the One-shot Protocol, here is a simple generic sample:
Let's say during training, you can train your model on the following actions:
at test time, in contrast, the following classes are contained with one reference sequence per class:
Thank you very much. I've got it.
在 2021-12-13 18:22:41,"Raphael Memmesheimer" @.***> 写道:
To elaborate on the One-shot Protocol, here is a simple generic sample:
Let's say during training, you can train your model on the following actions:
jumping running walking brushing teeth
at test time, in contrast, the following classes are contained with one reference sequence per class:
falling walking drinking
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.
I am still confused how the skeleton is related to the image representation. Could you please offer the related code?
I am also confused how the skeleton is related to the image representation. Could you please offer the related code?Thank you
@RCpengnan i just added a simple example in the representation
subfolder.
I'd like to know why the skeleton data of NTU 120 in your work are the series pictures rather than the skeleton data,such as the coordinate data. Are the pictures feature vectors? How do you get them?