3D Object Detection and Annotation
Φόρτωση...
Ημερομηνία
2021-07-01
Συγγραφείς
Stamatoudis, Ioannis
Τίτλος Εφημερίδας
Περιοδικό ISSN
Τίτλος τόμου
Εκδότης
Δικαιώματα
Default License
Άδειες
Παραπομπή
Παραπομπή
Περίληψη
Περίληψη
The area of deep learning with 3d data is underexplored due to the fact that only lately 3d
datasets become available. One type of 3d data is point clouds that usually are created by lidar
sensors. There is intensive research, in the recent years, on the field of object detection and
annotation with 3d data and many different state-of-the-art methods have been presented.
The purpose behind this thesis, is to try to improve one of these methods, by combining it
with parts of different methods or use it in parallel with another method. The motivation
behind this thesis, is to research if it is possible to get better results by modifying a method.
An application of this research is that there could be created a guideline for improving an
already successful method.
The selected method in 3d object detection and annotation to be modified was Pointpillars.
One part of the exploration was to substitute part of this method with parts from different
methods and test its performance. The selected parts to be added were residual networks and
inception networks, that were successful in 2d object recognition. Pointpillars, first creates a
pseudo-image from 3d data and then uses 2d convolutions to recognize patterns, and at this
stage the above parts can be introduced.
The other part of the exploration is to train parallelly with another method. The selected
method was Complex YOLO, and the recognized features of each method were combined and
then it was performed detection of objects. The goal is to check if the two parallel methods
can outperform the original Pointpillars method.
The results of the experiments showed that there was not improvement in the results both with
the modifications and the parallel training. Except that, some of the modified models were
more computationally expensive and inefficient compared to the original method.
Περιγραφή
Λέξεις-κλειδιά
3D object detection, 3D object annotation