The original paper is in English. Non-English content has been machine-translated and may contain typographical errors or mistranslations. ex. Some numerals are expressed as "XNUMX".
Copyrights notice
The original paper is in English. Non-English content has been machine-translated and may contain typographical errors or mistranslations. Copyrights notice
Nous présentons un système efficace pour intégrer des modules de classification générative à tir nul dans un détecteur dense de type YOLO afin de détecter de nouveaux objets. La plupart des nouvelles méthodes de détection d'objets à double étape sont obtenues en affinant la branche de sortie de classification mais ne peuvent pas être appliquées à un détecteur dense. Notre système utilise deux voies pour injecter des connaissances sur de nouveaux objets dans un détecteur dense. La première consiste à injecter la confiance de classe pour de nouvelles classes à partir d'un classificateur formé sur des données synthétisées via un générateur en deux étapes. Ce générateur apprend une fonction de mappage entre deux espaces de fonctionnalités, ce qui entraîne de meilleures performances de classification. La deuxième voie consiste à recycler la tête du détecteur avec des cartes de caractéristiques synthétisées à différents niveaux d'intensité. Cette approche augmente considérablement l’objectivité prédite pour les nouveaux objets, ce qui constitue un défi majeur pour un détecteur dense. Nous introduisons également un mécanisme d'arrêt et de rechargement pendant le recyclage pour l'optimisation entre les couches principales afin de mieux apprendre les fonctionnalités synthétisées. Notre méthode assouplit la contrainte sur l'architecture de la tête de détection dans la méthode précédente et a nettement amélioré les performances sur l'ensemble de données MSCOCO.
KuanChao CHU
University of Tokyo
Hideki NAKAYAMA
University of Tokyo
The copyright of the original papers published on this site belongs to IEICE. Unauthorized use of the original or translated papers is prohibited. See IEICE Provisions on Copyright for details.
Copier
KuanChao CHU, Hideki NAKAYAMA, "Two-Path Object Knowledge Injection for Detecting Novel Objects With Single-Stage Dense Detector" in IEICE TRANSACTIONS on Information,
vol. E106-D, no. 11, pp. 1868-1880, November 2023, doi: 10.1587/transinf.2022EDP7216.
Abstract: We present an effective system for integrating generative zero-shot classification modules into a YOLO-like dense detector to detect novel objects. Most double-stage-based novel object detection methods are achieved by refining the classification output branch but cannot be applied to a dense detector. Our system utilizes two paths to inject knowledge of novel objects into a dense detector. One involves injecting the class confidence for novel classes from a classifier trained on data synthesized via a dual-step generator. This generator learns a mapping function between two feature spaces, resulting in better classification performance. The second path involves re-training the detector head with feature maps synthesized on different intensity levels. This approach significantly increases the predicted objectness for novel objects, which is a major challenge for a dense detector. We also introduce a stop-and-reload mechanism during re-training for optimizing across head layers to better learn synthesized features. Our method relaxes the constraint on the detector head architecture in the previous method and has markedly enhanced performance on the MSCOCO dataset.
URL: https://global.ieice.org/en_transactions/information/10.1587/transinf.2022EDP7216/_p
Copier
@ARTICLE{e106-d_11_1868,
author={KuanChao CHU, Hideki NAKAYAMA, },
journal={IEICE TRANSACTIONS on Information},
title={Two-Path Object Knowledge Injection for Detecting Novel Objects With Single-Stage Dense Detector},
year={2023},
volume={E106-D},
number={11},
pages={1868-1880},
abstract={We present an effective system for integrating generative zero-shot classification modules into a YOLO-like dense detector to detect novel objects. Most double-stage-based novel object detection methods are achieved by refining the classification output branch but cannot be applied to a dense detector. Our system utilizes two paths to inject knowledge of novel objects into a dense detector. One involves injecting the class confidence for novel classes from a classifier trained on data synthesized via a dual-step generator. This generator learns a mapping function between two feature spaces, resulting in better classification performance. The second path involves re-training the detector head with feature maps synthesized on different intensity levels. This approach significantly increases the predicted objectness for novel objects, which is a major challenge for a dense detector. We also introduce a stop-and-reload mechanism during re-training for optimizing across head layers to better learn synthesized features. Our method relaxes the constraint on the detector head architecture in the previous method and has markedly enhanced performance on the MSCOCO dataset.},
keywords={},
doi={10.1587/transinf.2022EDP7216},
ISSN={1745-1361},
month={November},}
Copier
TY - JOUR
TI - Two-Path Object Knowledge Injection for Detecting Novel Objects With Single-Stage Dense Detector
T2 - IEICE TRANSACTIONS on Information
SP - 1868
EP - 1880
AU - KuanChao CHU
AU - Hideki NAKAYAMA
PY - 2023
DO - 10.1587/transinf.2022EDP7216
JO - IEICE TRANSACTIONS on Information
SN - 1745-1361
VL - E106-D
IS - 11
JA - IEICE TRANSACTIONS on Information
Y1 - November 2023
AB - We present an effective system for integrating generative zero-shot classification modules into a YOLO-like dense detector to detect novel objects. Most double-stage-based novel object detection methods are achieved by refining the classification output branch but cannot be applied to a dense detector. Our system utilizes two paths to inject knowledge of novel objects into a dense detector. One involves injecting the class confidence for novel classes from a classifier trained on data synthesized via a dual-step generator. This generator learns a mapping function between two feature spaces, resulting in better classification performance. The second path involves re-training the detector head with feature maps synthesized on different intensity levels. This approach significantly increases the predicted objectness for novel objects, which is a major challenge for a dense detector. We also introduce a stop-and-reload mechanism during re-training for optimizing across head layers to better learn synthesized features. Our method relaxes the constraint on the detector head architecture in the previous method and has markedly enhanced performance on the MSCOCO dataset.
ER -