Synthetic Dataset Generation using Nvidia Omniverse Replicator
In this work, which is the second part of my thesis; Enhanced Object Recognition in Hemispherical Images using Data Augmentation and Synthetic Image Integration, I generate 2D scenes from 3D scenes using omniverse replicator embedded in the Nvidia Omniverse Code. These scenes contain 13 iconic underepresented and underperforming classes out of the 365 classes in Objects365 dataset. The classes are Apple, Pomegranate, Papaya, Antelope, parrot, seal, butterfly, donkey, game board, monkey, rabbit, comb, and table tennis paddle. The generated scenes contain objects with their respective 2D bounding box information. Although some conversion and processing has to be made to make the bounding box information of the objects in the images in npy YOLO compatible. Here is the link to the processed dataset, arranged in train, val, images and labels in YOLO format.
Here is the link to the raw, unprocessed synthetic dataset. It is about 14GB.
The processed synthetic dataset is then used to finetune and improve the performance of a YOLOv7 model trained on 5.1 million images generated from Objects365 using data augmentation techniques (first part of my thesis). This 5.1 million images contain several type of images ranging from perspective to panoramic, wide-angle lenses and hemispherical cameras. Performance of the YOLov7 model on 8 out of these 13 classes improved with the least being 21% and maximum being 183%
All 3D assets used in this workk are obtained for free from the internet.
The python sripts alongside the samples of 3D assets used are provided.