University of Twente Student Theses


Diffuse more objects with fewer labels

Heuvel, L.P.W. van den (2023) Diffuse more objects with fewer labels.

[img] PDF
Abstract:Training object detection models requires a lot of labelled images that are often manually annotated by humans. Therefore, training object detection models with fewer labelled images is important. This work adapts a domain adaptation strategy, where an object detection model is trained on a source domain, and evaluated on the target domain where the labelled images are sparse. This work introduces a method to improve the detection performance in target domain without any fine-tuning. Using DiffusionDet a diffusion model for object detection, this work proposes a method to improve the performance of the detections in the target domain without any finetuning in trade for extra runtime. These improved detections are used as pseudo-labels to fine-tune the model in out-domain dataset of images. Using a human in the loop, the pseudo-labels were verified to make sure that the quality of the labels was sufficient. Using this method, the human effort required to improve an object detection model can be reduced compared against training with fully annotated samples only.
Item Type:Essay (Master)
TNO, The Hague
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Interaction Technology MSc (60030)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page