search

open-mmlab / mmskeleton

A OpenMMLAB toolbox for human pose estimation, skeleton-based action recognition, and action synthesis.
Apache License 2.0
2.94k stars 1.04k forks source link

How to test one by one video? #195

Open JasOlean opened 5 years ago

JasOlean commented 5 years ago

Code version (Git Hash) and PyTorch version

Dataset used

Expected behavior

Actual behavior

Steps to reproduce the behavior

Other comments

After obtaining weights file, how to test one by one video file?

daylightsaving2017 commented 5 years ago

Due to unknown tricks, I cannot test the video using the for loop (because it would always return the inference results of the first video). Hence, I solve this problem by calling the command line inside the python script with different input arguments iteratively.

daylightsaving2017 commented 5 years ago

@JasOlean

JasOlean commented 5 years ago

Where can I reference it?

daylightsaving2017 commented 5 years ago

@JasOlean Hahhh. Meet you again. I suggest that you could first use openpose to extract the skeletons from all your test videos with storing the results in the 'data/openpose_estimation' directory and then use command line functions imported from the os package installed in the python environment. About how to implement the virtual input, you could use some string concat functions. This idea is also inspired by the author of this project in the openpose calling mechanism which has been provided with the source code.

JasOlean commented 5 years ago

Yes. I already done all that steps. I even trained my dataset. But, accuracy is about over 41 %. That is why, I want to know what I need to do to obtain better results.

JasOlean commented 5 years ago

When I test one by one video, I got memory dump error.

daylightsaving2017 commented 5 years ago

@JasOlean About the accuracy problem, I advise you to read the corresponding paper more carefully. In the paper, the author wrote a section to explain the low accuray in the object/background-related video. Hence, this is not the fault or shortcome of the algorithm but the problem of inappropriate application scenes. If you wanna get significantly better recognition results on the current mainstream datasets, you should focus on the CNN(even with optical flow) methods like TSN/ C3D algorithms.

daylightsaving2017 commented 5 years ago

@JasOlean About the memory dump error, I am not sure what really invoked the problems by such limited information. However, I suggest you use a GPU with not lower than 8GB display memory.

JasOlean commented 5 years ago

Thank you so much for your advice. I will find it. so, for testing, how to test particular video after getting weights like your demo video. https://github.com/yysijie/st-gcn/blob/master/resource/info/demo_video.gif

yysijie commented 5 years ago

@JasOlean, @daylightsaving2017, The latest Openpose provides Python-API. It let the end-to-end inference from video become possible, which may help you in your application.
See the new demo script mentioned in the README.MD.