Skip to content

Latest commit

 

History

History
39 lines (26 loc) · 1.72 KB

README.md

File metadata and controls

39 lines (26 loc) · 1.72 KB
Raw

Interestingness_ICMR

Code for the work: Deep Pairwise Classification & Ranking for Predicting Media Interestingness by Jayneel Parekh, Harshvardhan Tibrewal, Sanjeel Parekh

Requirements

  • numpy
  • tensorflow >= 1.3.0
  • keras >= 2.1.3

Data

Contains code for training/testing our system on the MediaEval Predicting Media Interestingness Dataset (Link for task description: http://www.multimediaeval.org/mediaeval2017/mediainterestingness/index.html).

The Dataset needs to be acquired by the user himself/herself.

The Network Weights file (for testing system we trained) exceeds the max file size of git, thus not uploaded here. Weights for image subtask: https://drive.google.com/file/d/13prsjQ5FwsfHrPjkw-dXJorq6C53hd5l/view?usp=sharing

Usage

Command to train image interestingness system:

python train_nn.py [data_type] [operation] [ranker]

Functionality currently available to user is 

(a) data_type : Selects whether to run experiments on images or videos. (Options : [image , video] )

(b) operation : Training or Testing or making predictions for any set of images/video-shots if
    provided with appropriate feature representation files for images/video (respresented as feature vector of size 4096)
    (Options : [train ,  test] )
    
(c) ranker: selection of ranking algorithm for prediction (Options: [mih_to, mih_ro, sp, pp] )

Example : python train_nn.py image train mih_to

Currently the model is fixed and uses fc7 features (AlexNet) for images and c3d features for videos.

Next possible update is to allow more flexibility in feature respresentation of the input.