Class Incremental Learning For Video Action Classification

paper information

Title: CLASS INCREMENTAL LEARNING FOR VIDEO ACTION CLASSIFICATION (Accepted by icip 2021)
Authors: Jiawei Ma, Xiaoyu Tao, Jianxing Ma, Xiaopeng Hong, Yihong Gong

Introduction

This paper presents a CIL method for video action classification. This method uses CNN to extract features for each frame, and combines the advantages of GWR that can automatically grow or shrink to model the feature manifold for each action class. The Knowledge Consolidation (KC) method is introduced to alleviate forgetting by separating the feature manifolds of old class and new class.

Code Environment

• ubuntu 16.04
• python 3.6
• numpy 1.19
• pytorch 1.3

prepare

• Download the project and action datasets.

  The download url is dataset. Extract password is fffv

• Activate your environment

• Modify the absolute path in the project

train and test for each incremental session

cd ./CIL-for-video-action-recognition

python train.py

Contact Email

If you have questions, please contact the Email: [email protected]