search

ruiyan1995 / Group-Activity-Recognition

A novel Participation-Contributed Temporal Dynamic Model for Group Activity Recognition
25 stars 6 forks source link

Can't run `python GAR.py` . #2

Closed hahanigehaha233 closed 5 years ago

hahanigehaha233 commented 5 years ago

There is some confidence in your update code, Please push the runnable code version. Thanks.

hahanigehaha233 commented 5 years ago

PS: the problem is didn't download VD dataset.

ruiyan1995 commented 5 years ago

You need to download the dataset manually in step ZERO. And I will clean some error in the update code.

hahanigehaha233 commented 5 years ago

Thanks

hahanigehaha233 commented 5 years ago

How many feature_size during test CAD data_set in action_level.py I can set.

ruiyan1995 commented 5 years ago

How many feature_size during test CAD data_set in action_level.py I can set.

As same as VD! The details are as follow,

In experiments on Collective Activity Dataset, 10 time-steps and 3000 hidden nodes are used for the Single-Person LSTM and a softmax layer is deployed for the classification in the “One” Network. The number of sub-group is set to Nд = 1, namely, we do not need to divide the group. In “One to Key” Network (OKN), Interaction Bi-LSTM has a five time-steps (there are five individuals in one sub-group) and 1000 nodes; and Aggregation LSTM has five time-steps and 1000 nodes. Since the number of individuals in this dataset is varying from 1 to 12. We select five effective persons for each frame and regard them as an entire group. If the number of persons is less than five, we take a full-zero matrix as the tracklets of new person.