Batch-Based Activity Recognition From Egocentric Photo-Streams

Introduction

This repository contains the code used in the paper Batch-Based Activity Recognition From Egocentric Photo-Streams originally presented at the EPIC workshop @ ICCV 2017

If you use this code, please consider citing:

  @InProceedings{Cartas_2017_ICCV,
     author = {Cartas, Alejandro and Dimiccoli, Mariella and Radeva, Petia},
     title = {Batch-Based Activity Recognition From Egocentric Photo-Streams},
     booktitle = {The IEEE International Conference on Computer Vision (ICCV) Workshops},
     month = {Oct},
     year = {2017}
  }

Contents

1. Installation
2. Usage
3. Downloads

Installation

1. Clone this repository

   git clone --recursive https://github.com/gorayni/iccv_epic_2017.git
   

2. Download the NTCIR-12 dataset at http://ntcir-lifelog.computing.dcu.ie/NTCIR12/.

3. Create a symbolic link datasets/ntcir/images pointing to NTCIR_Lifelog_formal_run_Dataset/NTCIR-Lifelog_formal_run_images.

4. Split the data by executing the Creating data splits for VGG-16 training notebook. This will create the cross-validation splits in the directory data.

Training

Training the VGG-16+LSTM:

timestep=5
./run_training_vgg16+lstm.sh $timestep

Downloads

1. VGG-16 [weights]
2. VGG-16+LSTM [Timestep 5] [Timestep 10]
3. Piggyback VGG-16+LSTM [Timestep 5 Overlap 2] [Timestep 10 Overlap 3]