Siamski sledilnik s segmentacijo za robustno lokalizacijo tarče

Karamatić, Boris (author), Kristan, Matej (mentor)

Na področju sledilnikov trenutno prevladujeta dve metodi, in sicer sledenje z diskriminativnimi korelacijskimi filtri in sledenje s siamskimi mrežami. Obe metodi temeljita na sledenju s pravokotnimi regijami in imata probleme s sledenjem objektov, ki so slabo aproksimirani z očrtanimi pravokotniki. V magistrskem delu smo predlagali nov sledilnik SiamDGC, ki je združil dobre lastnosti sledilnikov s pravokotnimi regijami in dobre lastnosti segmentacijskih modelov za natančno segmentacijo. Sledilnik SiamDGC deluje tako, da na območju lokalizacije iz SiamFC izreže sliko, doda mapo evklidske razdalje in mapo odziva iz SiamFC ter izvede segmentacijo objekta sledenja. Na osnovi segmentacijske maske se nato posodobi lokalizacija v SiamFC za iskanje v sledečih sličicah. Sledilnik SiamDGC smo testirali in primerjali s sledilnikom SiamFC na podatkovnih zbirkah VOT2016 in DAVIS2016. Pri tem smo gledali natančnost in robustnost sledilnika. Ugotovili smo, da je za zanesljivo sledenje posodabljanje predloge sledenja zelo pomembno. Dodatek segmentacijske metode izboljša natančnost sledilnika SiamFC s posodabljanjem predloge na zbirki VOT2016 za 11,82%/6,26% in robustnost za 0,00%/8,06% pri primerjavi mask/očrtanih pravokotnikov. Na zbirki DAVIS2016 pa se natančnost izboljša za 27,20%/7,59% in robustnost za 75,00%/75,00% pri primerjavi mask/očrtanih pravokotnikov.

• Language: Slovenian
• Keywords: sledilnik, segmentacija, mreže, sledenje, detekcija, lokalizacija
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2020
• PID: 20.500.12556/RUL-122050
• COBISS.SI-ID: 40148227
• Publication date in RUL: 18.11.2020

Secondary language

• Language: English
• Title: Siamese tracker with segmentation for robust target localization

Abstract

Two methods dominate in the field single-target visual object tracking. Tracking with discriminant correlation filters and tracking with Siamese networks. Both methods are based on tracking with rectangular regions and have difficulty tracking objects that are poorly approximated with a bounding box. In the master's thesis we proposed a new SiamDGC tracker which combines good properties from the trackers with rectangular regions and of the segmentations models with accurate segmentations. The SiamDGC tracker works by cutting out an image from the localization returned by the SiamFC, adding an Euclidean distance map and a response map from the SiamFC and then segmenting the tracked object. Based on the segmentation mask we then update the localization and search area for SiamFC in the following images. The SiamDGC tracker was compared with the SiamFC tracker on the VOT2016 and DAVIS2016 databases in terms of accuracy and robustness. We concluded that updating the tracking template online is important. Addition of the segmentation method improved the accuracy of the SiamFC tracker, with template updating, on the VOT2016 database by 11.82%/6.26% and robustness by 0.00%/8.06% when comparing masks/bounding boxes. On the DAVIS2016 database accuracy is improved by 27.20%/7.59% and robustness by 75.00%/75.00% when comparing masks/bounding boxes.

• Keywords: tracker, segmentation, networks, tracking, detection, localization