Naslov
Comparison of Synthetic vs Augmented Dataset Metrics on YOLO Algorithm

Autori
Grebo, Alen ; Sokol, Domina ; Wolf, Josip, Gašparović, Goran

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, znanstveni

Izvornik
Proceedings of 20th International Conference "Aviation and Cosmonautics" (AviaSpace-2021) / - , 2022

Skup
20th International Conference "Aviation and Cosmonautics" (AviaSpace-2021)

Mjesto i datum
Moskva, Ruska Federacija, 22.11.2021. - 26.11.2021

Vrsta sudjelovanja
Predavanje

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
Synthetic datasets, YOLO, Augmented datasets

Sažetak
In this paper a step by step guide for synthetic data generation will be presented. This can prove effective when training algorithms like You Only Look Once (YOLO), YOLO v2, YOLO v3 as they are one-stage detection algorithms. When taking into consideration how time expensive it is to prepare real world data versus synthetic data, synthetic data wins by a large margin. In this case, it was necessary to produce a synthetic dataset of specific targets on the ground taken from an unmanned aerial vehicle (UAV) from several flight heights or levels. Producing a synthetic dataset makes it possible to train the model for a highly specific tasks, i.e. specific target or object class. The first step was determining ground sample distance (GSD), or how much of the ground is covered with a single image frame taken by the UAV. In the case at question, it was necessary to produce synthetic images from two flight levels (152 m and 54 m AGL) with varying shape, colour, and alpha numeric characters. Image generation was split into two parts, first part being determining the random position of the shape in the background image, and the second one, random shape generation with randomly chosen alpha numeric characters and colours. From ground sample distance, the information on how large (or small) objects appear to be in a single image frame was obtained, providing the bounding boxes for the objects that needed to be filled with random shapes. This was done for two flight levels and for two object sizes, yielding four bounding boxes. This proved to be a success as training labels for YOLO algorithm come in a form of a data vector with four numeric values and one string value. Four numeric values correspond to x and y values of a scribed rectangle while other two are width and height. String values is the target class. This allowed the generation of labelled synthetic images ready for the training of said algorithm. Total output of this approach is 134768 images with labels ready for training for only one class, in this case, named target. YOLO v2 was trained on this dataset and achieved an average precision of 64% which was significantly better then in the case of real-world drone data, only 31%. Future work will be focused on splitting the dataset into nine classes or nine targets and improving the average precision.

Izvorni jezik
Engleski

Znanstvena područja
Računarstvo, Strojarstvo, Zrakoplovstvo, raketna i svemirska tehnika, Interdisciplinarne tehničke znanosti

POVEZANOST RADA