University of Twente Student Theses

Login

Generating Sythetic Training Images for Instance Segmentation using Salient Object Detection and Image Compositions

Naik, Pratik (2021) Generating Sythetic Training Images for Instance Segmentation using Salient Object Detection and Image Compositions.

[img] PDF
68MB
Abstract:Deep learning in computer vision research requires a lot of data. This data is usually manually annotated. This manual annotation process is expensive and time-consuming. Manually annotating is difficult for an individual researcher or a small research group because there is no way to determine the number of images needed to sufficiently train a model. To solve this problem, a synthetic image generation pipeline is proposed and tested for the instance segmentation task with Mask R-CNN models. In this study, synthetic images and their annotations are generated using foreground extraction and image compositing. A salient object detection network called U^2-Net is used for the foreground extraction step. Images are composed with the extracted foregrounds using operations like random flipping, scaling, and rotating. Along with this, the effects of adding noise, adding unlabelled instances are studied. The effects of using hybrid datasets and initializing training with synthetic data and then retraining the model with real image data are also studied. Generating synthetic image datasets required 20% of the time needed to manually annotate images. However, models trained on synthetic images only have 50% of the performance of the model trained on real image datasets.
Item Type:Essay (Master)
Faculty:ET: Engineering Technology
Subject:52 mechanical engineering, 54 computer science
Programme:Mechanical Engineering MSc (60439)
Link to this item:https://purl.utwente.nl/essays/87698
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page