Intermediate CNN Features

This repository contains the implementation of the feature extraction process described in Near-Duplicate Video Retrieval by Aggregating Intermediate CNN Layers. Given an input video, one frame per second is sampled and its visual descriptor is extracted from the activations of the intermediate convolution layers of a pre-trained Convolutional Neural Network. Then, the Maximum Activation of Convolutions (MAC) function is applied on the activation of each layer to generate a compact layer vector. Finally, the layer vector are concatenated to generate a single frame descriptor. The feature extraction process is depicted in the following figure.

Prerequisites

Python 2
Caffe or Tensorflow (CPU or GPU version)

Getting started

Installation

Clone this repo:

git clone https://github.com/MKLab-ITI/intermediate-cnn-features
cd intermediate-cnn-features

Install all dependencies:

pip install -r requirements.txt

Feature Extraction

Provide a file of videos or images to the corresponding argument. Each line of the given file have to contain the full path to one video or image.
Select one of the two supported frameworks to extract features: Caffe or Tensorflow.
Choose one of the three supported CNN architectures based on the selected framework. CNN architectures for each framework:
- Caffe: GoogleNet, VGG, ResNet
- Tensorflow: VGG, ResNet, Inception
Provide the required files for the pre-trained networks. Downloads for Caffe and Tensorflow.
Provide an output path. The generated files depends on the input file.
- Video list: a file is generated for each video with name <video_basename>_<selected_network>.npy.
- Image list: a file with name <selected_network>_features.npy is generated. Each vector in the 0 axis of the stored numpy array correspond to the image in the input file.
Caffe example:

python feature_extraction.py --video_list <video_file> --network googlenet --framework caffe --output_path test/ --prototxt bvlc_googlenet/deploy.prototxt --caffemodel bvlc_googlenet/bvlc_googlenet.caffemodel

Tensorflow example:

python feature_extraction.py --image_list <image_file> --network vgg --framework tensorflow --output_path test/ --tf_model slim/vgg_16.ckpt

Citation

If you use this code for your research, please cite our paper.

@inproceedings{kordopatis2017near,
  title={Near-Duplicate Video Retrieval by Aggregating Intermediate CNN Layers},
  author={Kordopatis-Zilos, Giorgos and Papadopoulos, Symeon and Patras, Ioannis and Kompatsiaris, Yiannis},
  booktitle={International Conference on Multimedia Modeling},
  year={2017}
}

Related Projects

NDVR-DML FIVR-200K

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details

Contact for further details about the project

Giorgos Kordopatis-Zilos (georgekordopatis@iti.gr)
Symeon Papadopoulos (papadop@iti.gr)

LifangHe / intermediate-cnn-features