Authors: Arvid Paulsson and Daníel Jóhannsson

Title: Developing and Evaluating an Object Detection Application for Real-World Data

Authors: Albin Nilsson and Andreas Anderberg

Supervisors: Anders Heyden, Maria Juhlin (IKEA) and Martin Tegner (IKEA)

Examiner: Kalle Åström

Abstract:

This thesis explores a pipeline for automated object identification and interactive dot
annotation in UGC images. The approach combines Grounding DINO for object
localization based on text prompts, CLIP for object classification, and EfficientSAM
for instance segmentation, followed by a custom dot placement step. The system performs
particularly well on images containing fewer and more distinct products, where
visual clutter is minimal. Evaluation of the full pipeline has presented challenges due
to annotation inconsistencies and varying object prominence, making quantitative assessment
complex. Despite these difficulties, the method shows promising accuracy in
localizing and labeling both frequent and infrequent items, demonstrating the viability
of multimodal AI techniques for fine-grained object detection and segmentation.
The findings have implications for scalable labeling of UGC images.