Object ReID in an office environment: An empirical study

Abstract

Object Re-identification (ReID) is a fundamental task in computer vision, enabling systems to recognize and track the same object across different frames and viewpoints, lighting conditions, and environmental contexts. In robotic applications, reliable object ReID is essential for enabling robots to maintain persistent identity of objects over time. While person and vehicle ReID have been extensively studied, object-level ReID remains unexplored. In this work, we present an empirical comparative study of state-of-the art representation learning algorithms - DINO, DINOv2, Triplet, I-JEPA, and CLIP, which are applied to object ReID in an office environment. We construct a custom office dataset, capturing diverse office objects. Each image is cropped using Grounding DINO for object detection. We extract embeddings for each object instance and perform ReID by computing cosine similarity. Performance is assessed by measuring whether the top-matching image corresponds to the correct object, using Mean Average Precision, Top-1 and Top-5 metrics.