A foundational research framework for real-world abandoned object detection: train-free baseline and a standardized benchmark

Abstract

Developing an Abandoned Object Detection (AOD) system for real-world deployment requires fair comparisons and evaluation strategies that reflect practical challenges. To this end, we propose a foundational research framework comprising (1) a train-free baseline system and (2) a standardized benchmark, both grounded in real-world conditions. The baseline integrates key components to address two challenges often overlooked in prior work: recovery from intermittent detection failures and disambiguation of abandonment near pedestrians. It employs a class-agnostic tracker to maintain abandoned object identities across occlusions and detection lapses, demonstrated on backpack, suitcase, and handbag in our baseline, using CLIP-SORT with CLIP embeddings for broad object generalization. An owner identification module infers abandonment status using ownership cues derived from spatial and temporal relations between static objects and pedestrians. These components are integrated into a modular, train-free architecture, serving as a reference point for deployment-oriented AOD research. The benchmark includes a dataset and evaluation protocol. We introduce Tough Object Abandonment Scenario for Testing (TOAST), an evaluation-focused dataset designed to capture real-world abandonment scene complexity with diverse human-object interactions. The accompanying dual-component evaluation protocol assesses not only video-level detection accuracy but also object-level temporal consistency, measuring the continuity and reliability of predictions over time. Together, the proposed baseline and benchmark provide a solid foundation for advancing real-world AOD systems.