1 of 12

HackTheNorth

2021

Thomas J. Hu Bhavini Rathod

Nick Pysklywec Andy Huang

2 of 12

Table of Contents

01

02

03

04

Challenge

Solution

Demo

Future Plans

3 of 12

Challenge

01

4 of 12

CANSOFCOM - Video Summarization

To develop automated video summarization tools that can detect objects and events of interest in a video and make this content searchable. The participants must develop an application that can ingest a long video (e.g. one hour of a person walking in a city) and produce a searchable summary of this video.

5 of 12

Solution

02

6 of 12

Technology We Used

Pipeline

Colab

YouTube API

=> Image Category

=> Scrape google images

=> Transfer Learn

Detectron2 based transfer learning on special ops relevant categories

Simulation of searching & query process.

React

Frontend.

Displays a search bar, videos, and video summary.

7 of 12

Markup We Proposed

Step 1

Step 3

Step 2

8 of 12

03

9 of 12

Use cases

  • Aids with sitreps & AARs
  • Uncover overlooked info
  • Flag unlawful conduct
  • Long term analyses for strategic decisions

10 of 12

Improvements

04

11 of 12

Improvements

Analyze Longer Videos

Smart sampling, where interesting moments are sampled at higher rates

Scalable & Performant

noSQL DB indexed on Salient keywords�cold & hot storage

Video Chapters

Highlighting the location of the searched object and the amount of time it appears in the video

Activity Recognition

Patrolling, digging, loading & unloading.

Direction of movement.

GPS

01

03

04

02

12 of 12

Thanks!

Bhavini Rathod

Andy Huang

Thomas J. Hu

Nick Pysklywec