i3dFeatureExtraction 0.3.1.1

This package helps extract i3D features with ResNet-50 backbone given a folder of videos

Hello, I am Hao Vy Phan. I have develop this package using ResNet-50 to convert a video into an extracted i3D features numpy file.

Update in this version: The function can receive directly the path to just 1 video, if multiplefiles is set False.

Overview

Input: a directory which store 1 or more videos.

Output: 1 or many .npy files (extracted i3D features). Each features file is shaped n/16 * 2048 where n is the number of frames in the video

If there is a problem installing or implementing this package, please do not hesitate to contact me via my email. I am pleased to have people use my product.

---

Usage

Installation

Before installing my package, please install these pakages: * Opencv-Python==4.5.5

* torch==1.10.1+cu113

* torchaudio==0.10.1+cu113

* torchvision==0.11.2+cu113

* Or find your own python OS version of torch from this link: https://download.pytorch.org/whl/cu113/torch_stable.html

Installing them through pip install may raise errors. You can download the wheel files from the above links and run this code:

pip install torchvision-0.11.2+cu113-cp38-cp38-win_amd64.whl
pip install torchaudio-0.10.1+cu113-cp38-cp38-win_amd64.whl
pip install torch-1.10.1+cu113-cp38-cp38-win_amd64.whl
pip install opencv_python-4.5.5-cp38-cp38-win_amd64.whl

After 4 above packages, to install i3dFeatureExtraction package into your Python environment, run this code on your terminal:

pip install i3dFeatureExtraction

Implementing

The main function of this package is FeatureExtraction which converts a directory of videos into numpy feature files.

from i3dFeatureExtraction import FeatureExtraction
FeatureExtraction.generate(
    datasetpath="directory/of/input/videos",
    outputpath = "directory/to/store/output/numpy/files",
    pretrainedpath = "path/to/i3D/pretrained/weight",
    sample_mode = "oversample/center_crop"
    multiplefiles = True/False
)

• datasetpath (REQUIRED): path to videos.

• If multiplefiles is True (default), the datasetpath is the path to a directory which contains 1 or more videos.

• If multiplefiles is False, the datasetpath is the path to a video.

• outputpath (optional, default: “output/”): the proccessed numpy feature files would be stored in this directory.

• pretrainedpath (optional, default: “pretrained/”): the path of the pretrained i3d weight. If the weight is not existed, it will be downloaded and created manually.

• sample_mode (optional, default: “oversample/”) receive “oversample” or “center_crop”. If oversample, each frame will be cropped and flipped to be turned into 10 frames.

---

Structure of this package

I am not good at drawing UML diagram but I hope this image helps illustrate the package’s structure.

Credits

This code is based on the following repositories:

* pytorch-resnet3d

* pytorch-i3d-feature-extraction

* E2E-Action-Segmentation/feature_extraction/

I would like to extend a special thank-you to the original authors of these repositories for providing the foundation on which this implementation is built.