• norsk
    • English
  • norsk 
    • norsk
    • English
  • Logg inn
Vis innførsel 
  •   Hjem
  • Fakultet for informasjonsteknologi og elektroteknikk (IE)
  • Institutt for datateknologi og informatikk
  • Vis innførsel
  •   Hjem
  • Fakultet for informasjonsteknologi og elektroteknikk (IE)
  • Institutt for datateknologi og informatikk
  • Vis innførsel
JavaScript is disabled for your browser. Some features of this site may not work without it.

Adaptation of Deep Learning based large object detection models to Tiny Object Detection

Bhandari, Alen
Master thesis
Thumbnail
Permanent lenke
https://hdl.handle.net/11250/3003643
Utgivelsesdato
2022
Metadata
Vis full innførsel
Samlinger
  • Institutt for datateknologi og informatikk [5024]
Beskrivelse
Full text not available
Sammendrag
 
 
Despite achieving spectacular progress in deep learning (DL) for computer vision (CV) tasks such as object detection, DL models have yet not been able to show their performance when it comes to real-world usability, especially in natural environments. Especially, when it comes to tiny object detection, deep learning models could not perform well due to scale limitations imposed in these scenarios. It becomes even more challenging in cases with moving tiny objects such as birds, drones and others. The detection complexity increases due to their non-stationary position which requires DL models to detect tiny patterns in high-resolution images.

This Master's Thesis report tries to cover the Deep Learning aspect of tiny object detection. We experimentally evaluate the performance of state-of-the-art large object detection DL models on tiny object detection tasks. We train four models to test the feasibility of the large object detection models for tiny objects using publicly available high-resolution bird surveillance datasets, While we evaluate the feasibility of those models on a private dataset.

Based on the qualitative analysis, we realized that there are few deep learning models for bird detection in wind farms for high-resolution images. Due to this, our approach is a novel approach where our implemented methodology yields a result that outperforms the state-of-the-art model inference time by 1.8 sec/frame. Also, we beat the official benchmarks for the Mask R-CNN models benchmark by 7%.
 
Utgiver
NTNU

Kontakt oss | Gi tilbakemelding

Personvernerklæring
DSpace software copyright © 2002-2019  DuraSpace

Levert av  Unit
 

 

Bla i

Hele arkivetDelarkiv og samlingerUtgivelsesdatoForfattereTitlerEmneordDokumenttyperTidsskrifterDenne samlingenUtgivelsesdatoForfattereTitlerEmneordDokumenttyperTidsskrifter

Min side

Logg inn

Statistikk

Besøksstatistikk

Kontakt oss | Gi tilbakemelding

Personvernerklæring
DSpace software copyright © 2002-2019  DuraSpace

Levert av  Unit