Weakly Supervised Data Augmentation for LiDAR Based 3D Object Detection
Slabě supervisovaná příprava LiDARových dat pro detekci 3D objektů
Authors
Supervisors
Reviewers
Editors
Other contributors
Journal Title
Journal ISSN
Volume Title
Publisher
České vysoké učení technické v Praze
Czech Technical University in Prague
Czech Technical University in Prague
Date
Abstract
Navrhli jsme dvě lokální metody pro rozšíření datasetu: Vložení a Simulace pohybu. Vložení přidává do mračna bodů nové objekty. Tato metoda pomáhá nejvíce na nevyvážených datasetech, kde dokáže zvýšit zastoupení malých tříd. Simulace pohybu simuluje pozice všech pohybujících se objektů na základě jejich rychlosti a směru z 3D sledování. Pro obě tyto metody jsme navrhli algoritmus, který simuluje realistickou viditelnost. Pro Simulaci pohybu jsme navíc navrhli algoritmus, který zaplní tu část mračna bodů, která se odkryje posunutím objektů.
We propose two local 3D point cloud augmentations, Insertion and Movement simulation. Insertion method inserts objects bounding box to point cloud. This augmentation helps especially in case of an unbalanced dataset, as it increases number of exemplars for weakly represented classes. Movement simulation simulates positions of all moving objects in the future, based on their speed and direction from 3D tracking. For both of these augmentations, we design an algorithm that simulates realistic occlusion. For Movement simulation, we design an additional algorithm, which can fill parts of the scene that are uncovered by an objects movement.
We propose two local 3D point cloud augmentations, Insertion and Movement simulation. Insertion method inserts objects bounding box to point cloud. This augmentation helps especially in case of an unbalanced dataset, as it increases number of exemplars for weakly represented classes. Movement simulation simulates positions of all moving objects in the future, based on their speed and direction from 3D tracking. For both of these augmentations, we design an algorithm that simulates realistic occlusion. For Movement simulation, we design an additional algorithm, which can fill parts of the scene that are uncovered by an objects movement.