lunduniversity.lu.se

Digit@LTH

Faculty of Engineering, LTH

Denna sida på svenska This page in English

Digit@LTH: Events

MSc. by Lundström & Pettersson: 3D Privacy Masking using Unsupervised Monocular Depth Estimation

Output från modellerna och motsvarande resultat från 3D-maskningen.

Seminarium

From: 2022-06-14 09:00 to 09:45
Place: Seminar Room of Dept. of Automatic Control KC 3N27 and Zoom
Contact: anders [dot] robertsson [at] control [dot] lth [dot] se
Save event to your calendar


When: Tuesday June 14, 2022,  9.00-9.45

Where: The seminar will be held in the seminar room of the Department of Automatic Control, KC 3N27, Naturvetarvägen 18, Lund and Zoom:

 https://lu-se.zoom.us/j/65470487379?pwd=bkRWYzRSZEhuTlptdGRKNFhWTjlBQT09

Authors: Mattias Lundström & Jakob Pettersson
Advisors: Anders Robertsson, Dept. of Automatic Control, LTH
        Björn Olofsson, Dept. of Automatic Control, LTH     
        Joakim Ericson, Axis Communications

Examiner: Karl-Erik Årzén, Dept. of Automatic Control, LTH

 


3D Privacy Masking using Unsupervised Monocular Depth Estimation

Abstract

This thesis dives deeper within the area of monocular depth estimation, acquiring distance information from one single image. This is the result of dense prediction using deep CNN. This work contains a careful evaluation of  the performance of state-of-the-art depth estimation models, and investigates how they function and their limitations. This include depth metrics evaluation, where we compare the output of chosen models with ground truth data, temporal consistency evaluation between frames, depth range resolution evaluation, in order to get a sense of how far away the models are able to distinguish objects at different distances, and general analysis of scene- and image properties that affect the ability to estimate accurate depth maps. In addition, we have investigated how to improve the temporal consistency, utilising the information from previous frames in a video.

In addition to the work within depth estimation models, we have also implemented an algorithm we refer to as "3D Privacy Masking". Privacy masking is a typical task in camera surveillance, where a certain area of the image scene needs to be anonymised. We extend this further, by including depth information, like that from monocular depth estimation models, in the mask definition. This leads to a dynamic masking, where objects in front of the mask, as seen by the camera, are still visible. In other words. This masking technique leads to not losing information in the space between the camera and the privacy mask, as is the case for the regular 2D mask. We have proposed an end-to-end algorithm solution for this, as well as looking into a stabilising technique for a more robust masking result. Conclusively, this thesis strives to showcase the usability of monocular depth estimation in auxiliary computer vision tasks, like that of 3D privacy masking, outside the typical robotic domain, such as surveillance.