guiggh / hand_pose_action Goto Github PK

Hello, is there any link of data and pretriend model which can be downloaded.

Sorry to interrupt, I can not access the fist-person-hand-action dataset by your link:https://imperialcollegelondon.app.box.com/v/first-person-action-benchmark.
I have filled the form as required, but the website link just didn't work, can you fixed it?
Thanks.

Hi,

I noticed that a set of fixed camera extrinsics are provided, however, I am not sure how to retrieve the motion of the camera. Is there a way to retrieve camera extrinsics for every frame?

I found a comment on README.txt that the train-test ratio is based on this file, but how could I use it for splitting the train and test sets? Is each line on this file according to one image in the train set?

How to run it on the live webcam? Is it possible to test on our own images?

Hello, I found that the field of view of the depth map is wider than the field of view of the color map, what should I do to make them have the same field of view？

Hi. I've tried both load_example.m and load_example.py, however, I consider that for some sequences, the alignment from the skeleton and object_pose to image is problematic and does not globally match. I list some screenshots of frame examples below, which are generated by load_example.py; I also checked with the load_example.m, and the same issue exists:

Subject2 pour_milk sequence1 frame119

Subject3 pour milk sequence3 frame41

Subject6 pour milk sequence1 frame45

Subject4 read letter sequence3 frame88

Subject4 read letter sequence3 frame188

It seems that something related to the camera parameter goes wrong, and I here open the issue to seek help.

Furthermore, it seems that for Subject 5 open/close/pour liquid soap, I cannot find the object pose label for Sequence 1~4.

Many thanks for the valuable help!

Best regards,
Frances

Hi @guiggh ,

I am try to implement V2V PoseNet using your dataset. It requires centers for the dataset.

By any chance did you compute centers for Action dataset using deep prior++?

If please forward to my mail: [email protected]

Thanks!

Very impressive work! Thanks for sharing your dataset and providing a detailed codebase for us to start with. I am interested in an application that requires alignment of the color and depth images but it seems that they are shot in different coordinate frames. I am wondering if there is a way to map depth images to the color image frame?

I tried looking into the realsense github page, but the depth2image interface for that library is only for either bag files (raw realsense data file) or streaming data. It would be nice if you could point me to either a code base or something else.

I am unfamiliar with camera transform. I don't know if it is possible to do the mapping using the camera parameters you provided? Thanks!

Dear authors, thanks for your amazing work!

Would it be possible to provide a mapping from the indices in "action_sequences_normalized" to the corresponding action sequences? This would make it possible to use the image data of the sequences as well when training methods on your dataset.

Also, do you recall the coordinate system scale used for these normalized action sequences? I assume that for the unnormalized version, it's millimeters, but I can't find a good unit for the normalized ones. Meters don't seem to make much sense based on the coordinate ranges. Here's an example of a frame's normalized joint positions. Every 3rd coordinate corresponds to the same dimension, and dimension 2 (counting from 0) has a range of at least [0, 0.6] (check the 3rd and last element of the array), which does not correspond to a human hand's scale measured in meters.

Thank you very much in advance!

Hello,

Will the codes of the baseline LSTM for action recognition be open-source?

Best regards,
Yasser BOUTALEB

Hello, I want to know some questions about coordinate dimensions. How do you define the origin and direction of the coordinate system?

Will the implementation code be open-source?
Thanks for the reply.

Hi, I am planning to use the hand skeleton pose from action_sequenced_normalised. May I ask the post-processing of the data go through the same stages as if single images from skeleton.txt? Many thanks:)

I tried to download the dataset after filling the form, however, I come up with two problems:

1. When I tried to download all files into a single zip, it says file too large to download.
2. When I tried to download the video of a single subject, it always breaks after downloading about half of the data. (not network linking problem, just could not download file, and it says Fail, can not find the file.

Does anyone have the same problem?