xiongweiwu / actionnet

ActionNet_Tensorflow

This is Tensorflow implemented ActionNet which is used to recognize action from programming screencast.

Installation

This software has been tested on Ubuntu16.04(x64) using python3.5, opencv3.3, and tensorflow1.9 with cuda 9.0, cudnn9.0 and GTX-1080Ti GPU. Some other requirements:

• skimage
• imutils
• tqdm

Test model

Firstly, you should modify some parameters in 'analyze.py'.

• video_name: path to the video you want to test
• path: path to the dataset
• label_name: label.txt
• ck_path: path to Tensorflow check point
• num: class number before merge
• nums: class number after merge
• parts: T0 is the first frame and T1 is the second frame
• label_map: how to merge label Then run

Python3 analyze.py

The result will be stored in '/root/path/result' Here are some results:

Train model

1. Extract different region from video

python3 clip.py

You can also use python multiprocessing version

python3 clip_multiprocessing.py

The different region data will be stored in output_path 2. Merge all dataset Each video's different region data is stored in a floder. We need to merge all of them by

python3 merge_dataset.py

You can assign which folders will be merged by modifying dir_list. The merged result will be stored in 'output_path'. 3. Convert data to tf_record Modify the dataset path, this path would be the same with output_path in step 2.

python3 convert_data_to_tfrecord.py

tf_record would be stored in the current path 4. launch training Modify the dataset path in 'finetune.sh' Then run

./finetune.sh

Evaluate model

Modify the parameters in 'evaluation.py', this step is the same with Test model Then run

python3 evaluation.py

This step will show confusion matrix and some other measurment of this model.