Explanation of Testing about real-time-gesrec

[ahmetgunduz]

Hi @mhsattarian,
Thanks for kinds comments and well explained questions. You are right, I should have explained offline and online testing in more details.

what is the difference between run_online and run_offline scripts?

run online is for real-time analysis. It has different data loading methods as slide through the frames and make predictions in real time simulation environment. It runs online_test.py, in which the real-time detection and classification algorithms are implemented as it is described in the article.
run offline, on the other hand, is for training and evaluation of the models on the predefined clips of videos. This is the usual way of training a model. When you want to train a model for a dataset you need to train them first offline and then if you want to see their performance on a real-time scenario you would use run_online script by giving pretrained path, which is also depends either detector or classifier.

what changes should be applied to these scripts parameters in order to just test your model on a video?

As an example I am giving offline testing. But you can also do the same for online testing by providing two architecture paths (detector and classifier). SO the followings are for run_offline.sh

First you need to run offline_test.pyinstead of main.py, which is for training.
for dataset configurations (the followings are based on my path configurations for Egogesture dataset):

--root_path ~/ \
--video_path /data2/EgoGesture/images \
--annotation_path ~/Real-time-GesRec/annotation_EgoGesture/egogesturebinary.json\

for pretrained path configuration:

--resume_path MyRes3D-Ahmet/report/egogesture_resnetl_10_Depth_8_9939.pth \
-dataset egogesture \
--sample_duration 8 \
--model resnetl \
--model_depth 10 \
--batch_size 8 \
--n_classes 2 \ # if it is detector 2, if classifier # of classes in that dataset
--n_finetune_classes 2 \ # if it is detector 2, if classifier # of classes in that dataset
--modality Depth \

my pretrained model's names are in following convention <model_depth><sample_duration>.pth

why run_offline script calls model.py instead of offline_test.py?

I believe you meant main.py. main.py is for training, offline_test.py is for testing.

why there is no pretrain_path_det and pretrain_path_clf defined in run_online or run_offline scripts? aren't pretrained weights needed in order to test the model?

pretrain_path_{det or clf} is for detector and classifier architecture paths
and they are defined only for online testing. Because in online testing, it used detector and classifier together, However for offline training and testing you do this process separately, and you can only use pretrained_path in for those.

Hope this is clear. If not, you may check the paper again. As a last note, run offline is for training models and make them ready for online (real-time) testing.

Please let me know if this works for you.

[mhsattarian]

for example, can you answer these questions?:

• what is the difference between run_online and run_offline scripts?
• what changes should be applied to these scripts parameters in order to just test your model on a video?
• why run_offline script calls model.py instead of offline_test.py?
• why there is no pretrain_path_det and pretrain_path_clf defined in run_online or run_offline scripts? aren't pretrained weights needed in order to test the model?

[mhsattarian]

thank you @ahmetgunduz for answering all of the questions in detail.

[ahmetgunduz]

you are welcome! @mhsattarian