HDD dataset ?

[pgupta119]

Hi Team,
How I could use the same code for the HDD dataset?

[xumingze0308]

1. Download the dataset from HRI; 2) prepare the data structure as in README's data preparation; 3) add HDD to "data/data_info.json" (you can refer https://github.com/xumingze0308/TRN.pytorch/blob/master/data/data_info.json).

[pgupta119]

Thanks, @xumingze0308, one quick question about how to get RGB input and optical flow for the HDD dataset. I got the dataset but it follows the structure of the TRN paper (https://github.com/xumingze0308/TRN.pytorch). Do you have the RGB and optical flow inputs for HDD dataset.

[xumingze0308]

LSTR and TRN use the same data structure, so maybe you can directly use what you have.

I don't have the inputs for HDD. They are owned and maintained by Honda Research.

[tyh001]

请问如何使用私有数据集，如何转换成npy文件，或者说npy文件内容是什么

[xumingze0308]

Each numpy file contains the feature for one particular video and is in shape of "L x C", where L is the number of frames and C the feature dimension.

[tyh001]

Thank you. It seems that the script for video feature extraction is not provided in the code. Can you provide the script?

[xumingze0308]

Sorry, I don't think we have that plan.

[tyh001]

What is the internal contents of the target_perframe in the THUMOS'14 dataset? What is L x 22?

[xumingze0308]

L is the number of frames in a particular video. 22 is the number of classes, where the value is either 0 or 1 --- 0 means this action doesn't happen in this frame, and 1 means this action happens in this frame. Since THUMOS'14 is a multiclass dataset, so we cannot use one-hot to store the ground truth label.