TY - GEN

T1 - M3T

T2 - 37th International Conference on Image and Vision Computing New Zealand, IVCNZ 2022

AU - Khan, Mariia

AU - Abu-Khalaf, Jumana

AU - Suter, David

AU - Rosenhahn, Bodo

PY - 2023

Y1 - 2023

N2 - In this paper, we propose an extended multiple object tracking (MOT) task definition for embodied AI visual exploration research task - multi-class, multi-instance and multi-view object tracking (M3T). The aim of the proposed M3T task is to identify the unique number of objects in the environment, observed on the agent’s way, and visible from far or close view, from different angles or visible only partially. Classic MOT algorithms are not applicable for the M3T task, as they typically target moving single-class multiple object instances in one video and track objects, visible from only one angle or camera viewpoint. Thus, we present the M3T-Round algorithm designed for a simple scenario, where an agent takes 12 image frames, while rotating 360° from the initial position in a scene. We, first, detect each object in all image frames and then track objects (without any training), using cosine similarity metric for association of object tracks. The detector part of our M3T-Round algorithm is compatible with the baseline YOLOv4 algorithm [1] in terms of detection accuracy: a 5.26 point improvement in AP75. The tracker part of our M3T-Round algorithm shows a 4.6 point improvement in HOTA over GMOTv2 algorithm [2], a recent, high-performance tracking method. Moreover, we have collected a new challenging tracking dataset from AI2-Thor [3] simulator for training and evaluation of the proposed M3T-Round algorithm.

AB - In this paper, we propose an extended multiple object tracking (MOT) task definition for embodied AI visual exploration research task - multi-class, multi-instance and multi-view object tracking (M3T). The aim of the proposed M3T task is to identify the unique number of objects in the environment, observed on the agent’s way, and visible from far or close view, from different angles or visible only partially. Classic MOT algorithms are not applicable for the M3T task, as they typically target moving single-class multiple object instances in one video and track objects, visible from only one angle or camera viewpoint. Thus, we present the M3T-Round algorithm designed for a simple scenario, where an agent takes 12 image frames, while rotating 360° from the initial position in a scene. We, first, detect each object in all image frames and then track objects (without any training), using cosine similarity metric for association of object tracks. The detector part of our M3T-Round algorithm is compatible with the baseline YOLOv4 algorithm [1] in terms of detection accuracy: a 5.26 point improvement in AP75. The tracker part of our M3T-Round algorithm shows a 4.6 point improvement in HOTA over GMOTv2 algorithm [2], a recent, high-performance tracking method. Moreover, we have collected a new challenging tracking dataset from AI2-Thor [3] simulator for training and evaluation of the proposed M3T-Round algorithm.

KW - Embodied AI

KW - Multiple Object Tracking

KW - Scene Understanding

UR - http://www.scopus.com/inward/record.url?scp=85147999282&partnerID=8YFLogxK

U2 - 10.1007/978-3-031-25825-1_18

DO - 10.1007/978-3-031-25825-1_18

M3 - Conference contribution

AN - SCOPUS:85147999282

SN - 9783031258244

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 246

EP - 261

BT - Image and Vision Computing

A2 - Yan, Wei Qi

A2 - Nguyen, Minh

A2 - Stommel, Martin

PB - Springer Science and Business Media Deutschland GmbH

Y2 - 24 November 2022 through 25 November 2022

ER -