Skip to content

My experimentation around action recognition in videos. Contains Keras implementation for C3D network based on original paper "Learning Spatiotemporal Features with 3D Convolutional Networks", Tran et al. and it includes video processing pipelines coded using mPyPl package. Model is being benchmarked on popular UCF101 dataset and achieves result…

License

karolzak/conv3d-video-action-recognition

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

28 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

About

This repo contains Keras implementation for C3D network based on paper "Learning Spatiotemporal Features with 3D Convolutional Networks", Tran et al. and it includes video processing pipelines coded using amazing mPyPl package. Model is being benchmarked on UCF101 - a popular action recognition dataset - and achieves results similar to those reported by original authors.

Examples of action recognition tasks from UCF101:

ApplyEyeMakeUp Kayaking PlayingFlute

The whole beauty of C3D is that it's using Conv3D layers to learn spatiotemporal features from video frames and if trained on big enough amount of data it can be successfully used as a compact and uniform video feature extractor/descriptor. Features extracted from such model can easily be used to build a simple Linear SVM classifier. This approach achieves close-to-the-best results for most of the action recognition benchmark datasets while remaining very fast and efficient which is perfect for video processing tasks. C3D pretrained weights used in this exepriment are coming from training on Sports-1M dataset.


Prerequisites

This project heavily relies on mPyPl package and it's included in requirements.txt so in order to install it with all the other required dependencies you can run:

pip install -r requirements.txt

How to run it

End-to-end experiment

In order to run the experiment end-to-end (from original UCF101 videos, through preprocessing and features generation) you need to follow the steps below:

  1. Download UCF101 dataset and put videos with all classes subfolders under data/ucf101/videos/ so you have a structure such as this:

    data
    ├── ucf101
        ├── videos
            ├── ApplyEyeMakeup    
            │   ├── v_ApplyEyeMakeup_g01_c01.avi       
            │   ├── v_ApplyEyeMakeup_g01_c02.avi           
            │   └── ...  
            ├── ApplyLipstick             
            │   ├── v_ApplyLipstick_g01_c01.avi       
            │   ├── v_ApplyLipstick_g01_c02.avi            
            │   └── ...                
            └── ...
    
  2. Download Sports-1M pretrained C3D model and put it under models/

  3. Go to the notebook with end-to-end experiment for action recognition using pretrained C3D on UCF101 and execute it cell by cell


Final classification part of experiment only

In case you don't want to go through all the data loading and preprocessing steps (as it is very time and storage consuming) you can simply download feature vectors for each video and skip the data loading and preprocessing steps from end-to-end experiment:

  1. Download feature vectors for each video from UCF101 and put it under data/ucf101/videos/ so you have a structure such as this:

    data
    ├── ucf101
        ├── videos
            ├── ApplyEyeMakeup    
            │   ├── v_ApplyEyeMakeup_g01_c01.proc.c3d-avg.npy     
            │   ├── v_ApplyEyeMakeup_g01_c02.proc.c3d-avg.npy            
            │   └── ...  
            ├── ApplyLipstick             
            │   ├── v_ApplyLipstick_g01_c01.proc.c3d-avg.npy        
            │   ├── v_ApplyLipstick_g01_c02.proc.c3d-avg.npy              
            │   └── ...                
            └── ...
    
  2. Go to the notebook with end-to-end experiment for action recognition using pretrained C3D on UCF101 and execute it cell by cell skipping STEP 1 and STEP 2

About

My experimentation around action recognition in videos. Contains Keras implementation for C3D network based on original paper "Learning Spatiotemporal Features with 3D Convolutional Networks", Tran et al. and it includes video processing pipelines coded using mPyPl package. Model is being benchmarked on popular UCF101 dataset and achieves result…

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages