coranholmes/i3d_feature_extraction_resnet

I3D features extractor with resnet50 backbone

I3D_Feature_Extraction_resnet

This repo contains code to extract I3D features with resnet50 backbone given a folder of videos

This code can be used for the below paper. Use at your own risk since this is still untested.

• Weakly-supervised Video Anomaly Detection with Robust Temporal Feature Magnitude Learning

---

Credits

The main resnet code and others is collected from the following repositories.

• pytorch-resnet3d
• pytorch-i3d-feature-extraction

I modified and combined them and also added features to make it suitable for the given task.

Overview

This code takes a folder of videos as input and for each video it saves I3D feature numpy file of dimension 1*n/16*2048 where n is the no.of frames in the video

Usage

Pre-requisite

Install ffmpeg

sudo apt-get update
sudo apt-get install ffmpeg

Install requirements.txt

pip install -r requirements.txt

Setup

Download pretrained weights for I3D from the nonlocal repo

wget https://dl.fbaipublicfiles.com/video-nonlocal/i3d_baseline_32x2_IN_pretrain_400k.pkl -P pretrained/

Convert these weights from caffe2 to pytorch. This is just a simple renaming of the blobs to match the pytorch model.

python -m utils.convert_weights pretrained/i3d_baseline_32x2_IN_pretrain_400k.pkl pretrained/i3d_r50_kinetics.pth

Parameters

--datasetpath:       folder of input videos (contains videos or subdirectories of videos)
--outputpath:        folder of extracted features
--frequency:         how many frames between adjacent snippet
--batch_size:        batch size for snippets
--last_segment:	     whether to cut or pad last segment ("padding" or "cutting")

Run

python main.py --datasetpath=samplevideos/ --outputpath=output