Presentation

Authors: Michał Miotk, Kamil Orłow

Agenda

  1. Introduction
  2. Weapon Dataset Creator
  3. SynthDet
  4. ML->Dataset
  5. Future Work
  6. Links

Introduction

Generate weapons dataset from Unity engine instead taking photos of real guns on diffrent backgrounds or pictures from the Internet.

Article

Models 3D

Perception

  1. Dataset Creator
  2. Synthdet

Weapon DatasetCreator

rotate Glock with changing background
background 2D->3D setting objects, Raycast, object Pooling, 5FPS |_ |_ |_
Foreground
grid, Rotation 3^3 == 27(120degrees)
occlusion -> scaling (Background)
Output - 2FPS
SynthDet
Background - nalozone zdjecia na szesciany i walce, brak pelnych modeli
Foreground - moga byc obiekty z backgroundu na foreground - ustawia sie procentowa szanse
Occlusion - full implementation, ours is partial
Features - lights, blur, extra
ML - Datasets

1. EdgeAi Weapon Dataset

2. Dataset Generated in Unity

Based on free models from assets store - some of them has three instances

Categories

Detectron2 configuration

MODEL.WEIGHTS mask_rcnn_R_50_FPN_3x.yaml MODEL.BACKBONE.FREEZE_AT 2

Results

Precision on glock detection

trained on datesetgenerated with SynthDet

Results after trained on weapons dataset from edgeai

Results are presented in streamlit app

It needs own script to parse output

AWS sagemaker was used for test dataset

up to 10 labels, own scripts because no instance segmentation

Future Work

3D bounding box annotation

scanning przestrzeni 3D

Links

  • https://github.com/Unity-Technologies/SynthDet
  • https://github.com/Unity-Technologies/com.unity.perception
  • https://arxiv.org/abs/1902.09967
  • https://docs.unity3d.com/Manual/index.html
  • https://github.com/facebookresearch/detectron2
  • https://aws.amazon.com/sagemaker/
  • https://www.streamlit.io/
  • https://github.com/michalmiotk/gun
  • https://github.com/michalmiotk/WeaponDatasetCreator
  • https://blog.usejournal.com/worlds-first-synthetic-gun-detection-dataset-from-edgecase-ai-dbe3ea8eeb7e