search

happyharrycn / actionformer_release

Code release for ActionFormer (ECCV 2022)
MIT License
434 stars 77 forks source link

About THUMOS14 in this repo #33

Closed juipengchang closed 2 years ago

juipengchang commented 2 years ago

I have seen many different types of THUMOS14 in temporal action localization such as .json and .csv. Now I am making a new data set for temporal action localization, could you tell me if there any other method using the same type of data as yours? Thinks!

tzzcl commented 2 years ago

Hi, I'm not sure I fully understand your questions.

For THUMOS14, the dataset and annotations are deterministic. You may see difffent kinds of annotations, the original annotations are in the csv form. In our repo, to be compatible with ActivityNet, we just turn the csv format into the unified json format, no other changes are maken.

happyharrycn commented 2 years ago

To build on Chenlin's answer, we store video features in npz (numpy arrays) and video annotations in json format. Our json file follows the annotation format of ActivityNet-1.3. To adapt our features, you can write a dataloader to parse the json file and load the features, similar to our dataloader in this repo.

happyharrycn commented 2 years ago

Let me know if this issue is still open. Otherwise, I will mark it as resolved.

juipengchang commented 2 years ago

Not yet. Think you for reply!

tzzcl commented 2 years ago

Can you illustrate more about your questions?

juipengchang commented 2 years ago

Can you illustrate more about your questions?

I am trying to test many TAL models with my own data set. I have made my data set's label as your THUMOS14 label format and run your model successfully. However, other TAL models(such as CMCS) have other input format and label format(such as this). I don't know how to generate these files of my data set, so I tried to find some methods which can run with your THUMOS14 labels directly. Now I now that you use ActivityNet format, so I am trying to find some methods which can used on ActivityNet.

happyharrycn commented 2 years ago

The json file following ActivityNet format simply records the action annotations in the videos. You can write some Python code to convert another format into the json format and vice versa. Or you can re-write the dataloader to allow our code to interface with another format.

happyharrycn commented 2 years ago

I am closing this issue as the discussion seems irrelevant to the current repo. Send us an email if an offline discussion is needed.