Skip to content

kartikdutt18/CNN-LSTM-for-Activity-Recognition

Repository files navigation

CNN-LSTM for Activity Recognition

About the Dataset

UCF101 is one of the most popular action recognition datasets of realistic action videos. It consists of 13320 videos taken from YouTube, which are divided into 101 action categories. Each category contains videos between [100, 200]. UCF101 is comparatively more challenging dataset due to its large number of action categories from five major types: (1) human-object interaction, (2) body-motion only, (3) human-human interaction, (4) playing musical instruments, and (5) sports. Some categories have many actions such as sports, where most of the sports are played in a similar background, i.e., greenery.

Some of the videos are captured in different illuminations, poses, and from different viewpoints. One of the major challenges in this dataset is its realistic actions performed in real life, which is unique compared to other datasets where actions are performed by an actor. Advantage of such a network is that it can be trained on systems slower which don't have access to multiple GPU's. This model can achieve upto 78% accuracy with convnet as resnet.

To reproduce results: (1.) Train CNN for classification (2.) Extract Features (3.) Train LSTM

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages