Fully convolutional networks for semantic segmentation

Verifica disponibilità: Testo completo (sito dell'editore)

Abstract

Convolutional networks are powerful visual models that yield hierarchies of features. We show that convolutional networks by themselves, trained end-to-end, pixels-to-pixels, exceed the state-of-the-art in semantic segmentation. Our key insight is to build “fully convolutional” networks that take input of arbitrary size and produce correspondingly-sized output with efficient inference and learning. We define and detail the space of fully convolutional networks, explain their application to spatially dense prediction tasks, and draw connections to prior models. We adapt contemporary classification networks (AlexNet [20], the VGG net [31], and GoogLeNet [32]) into fully convolutional networks and transfer their learned representations by fine-tuning [3] to the segmentation task. We then define a skip architecture that combines semantic information from a deep, coarse layer with appearance information from a shallow, fine layer to produce accurate and detailed segmentations. Our fully convolutional network achieves state-of-the-art segmentation of PASCAL VOC (20% relative improvement to 62.2% mean IU on 2012), NYUDv2, and SIFT Flow, while inference takes less than one fifth of a second for a typical image.

Salvato in:
Autori: Jonathan Long (University of California, Berkeley); Evan Shelhamer iD (University of California, Berkeley); Trevor Jackson Darrell iD (University of California, Berkeley)
Pubblicato in: IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings, pp. 3431-3440
Data di pubblicazione:01/06/2015
Editore:Institute of Electrical and Electronics Engineers
Tipo di documento:Contributo a convegno
Lingua:inglese
DOI: 10.1109/cvpr.2015.7298965
ISSN:1063-6919, 2332-564X, 2575-7075
Accesso aperto: Nessuna copia ad accesso aperto nota: per le risorse in abbonamento usare Verifica disponibilità.
Citazioni ricevute:38.781
Soggetti:
Advanced Neural Network Applications (Physical Sciences > Computer Science > Computer Vision and Pattern Recognition)
Domain Adaptation and Few-Shot Learning (Physical Sciences > Computer Science > Artificial Intelligence)
Medical Image Segmentation Techniques (Physical Sciences > Computer Science > Computer Vision and Pattern Recognition)
fully convolutional network, semantic segmentation, skip connections, fine-tuning, transfer learning, AlexNet, VGG, GoogLeNet, inference efficiency, Pascal VOC, SIFT flow
Identificativi: OpenAlex W1903029394

Dati dell'indice centrale: OpenAlex (licenza CC0).

Riassunto:Convolutional networks are powerful visual models that yield hierarchies of features. We show that convolutional networks by themselves, trained end-to-end, pixels-to-pixels, exceed the state-of-the-art in semantic segmentation. Our key insight is to build “fully convolutional” networks that take input of arbitrary size and produce correspondingly-sized output with efficient inference and learning. We define and detail the space of fully convolutional networks, explain their application to spatially dense prediction tasks, and draw connections to prior models. We adapt contemporary classification networks (AlexNet [20], the VGG net [31], and GoogLeNet [32]) into fully convolutional networks and transfer their learned representations by fine-tuning [3] to the segmentation task. We then define a skip architecture that combines semantic information from a deep, coarse layer with appearance information from a shallow, fine layer to produce accurate and detailed segmentations. Our fully convolutional network achieves state-of-the-art segmentation of PASCAL VOC (20% relative improvement to 62.2% mean IU on 2012), NYUDv2, and SIFT Flow, while inference takes less than one fifth of a second for a typical image.
ISSN:1063-6919
2332-564X
2575-7075
DOI:10.1109/cvpr.2015.7298965