Failure Detection in Pedestrian Detection : A Fully Convolutional Neural Network approach
Author
Apostolopoulos, Christos
Term
4. term
Publication year
2016
Submitted on
2016-06-02
Pages
85
Abstract
At finde fodgængere i billeder er et centralt problem i computer vision med anvendelser, der kan påvirke hverdagen. Dette projekt forsøger at rette fejl i eksisterende fodgængerdetektorer ved at genvurdere deres output med et fuldt konvolutionelt neuralt netværk (FCN), en dybdelæringsmodel der behandler billeder på pixelniveau. Netværket blev trænet på flere datasæt, herunder et eget datasæt med delvist skjulte fodgængere (personer der er delvist dækket af objekter). Vi evaluerede ydeevnen ved at følge tabsværdien under træningen (iterationer) og ved at måle den gennemsnitlige Intersection over Union (IoU) på tværs af klasser, som angiver hvor godt de forudsagte områder overlapper de rigtige. Som detektor blev nøjagtigheden vurderet ved at sammenligne FCN'et, når det får detektorens output, med førende metoder og ved at sammenligne de raffinerede resultater med den oprindelige detektor uden FCN. Resultaterne varierer, men metoden viser lovende potentiale.
Detecting pedestrians in images is a core problem in computer vision with applications that can affect everyday life. This project aims to correct errors made by existing pedestrian detectors by re-evaluating their outputs with a fully convolutional neural network (FCN), a deep learning model that processes images at the pixel level. The network was trained on several datasets, including a custom dataset of partially occluded pedestrians (people partly hidden by objects). We evaluated performance by tracking training loss over iterations and by measuring mean Intersection over Union (IoU) across classes, which quantifies how well predicted regions overlap with the ground truth. As a detector, accuracy was assessed by comparing the FCN, when given detector outputs, with state-of-the-art methods and by comparing the refined results with the original detector without the FCN. Results vary across tests, but the approach shows promising potential.
[This abstract was generated with the help of AI]
Documents