Keypoint detection algorithms are typically based on handcrafted combinations of derivative operations implemented with standard image filtering approaches. The early layers of Convolutional Neural Networks (CNNs) for image classification, whose implementation is nowadays often available within optimized hardware units, are characterized by a similar architecture. Therefore, the exploration of CNNs for keypoint detection is a promising avenue to obtain a low-latency implementation, also enabling to effectively move the computational cost of the detection to dedicated Neural Network processing units. This paper proposes a methodology for effective keypoint detection by means of an efficient CNN characterized by a compact three-layer architecture. A novel training procedure is proposed for learning values of the network parameters which allow for an approximation of the response of handcrafted detectors, showing that the proposed architecture is able to obtain results comparable with the state of the art. The capability of emulating different detectors allows to deploy a variety of algorithms to dedicated hardware by simply retraining the network. A sensor-based FPGA implementation of the introduced CNN architecture is presented, allowing latency smaller than 1[ms].

Paolo Di Febbo, C.D.M. (2018). KCNN: Extremely-Efficient Hardware Keypoint Detection With a Compact Convolutional Neural Network. IEEE/CVF.

KCNN: Extremely-Efficient Hardware Keypoint Detection With a Compact Convolutional Neural Network

Stefano Mattoccia
2018

Abstract

Keypoint detection algorithms are typically based on handcrafted combinations of derivative operations implemented with standard image filtering approaches. The early layers of Convolutional Neural Networks (CNNs) for image classification, whose implementation is nowadays often available within optimized hardware units, are characterized by a similar architecture. Therefore, the exploration of CNNs for keypoint detection is a promising avenue to obtain a low-latency implementation, also enabling to effectively move the computational cost of the detection to dedicated Neural Network processing units. This paper proposes a methodology for effective keypoint detection by means of an efficient CNN characterized by a compact three-layer architecture. A novel training procedure is proposed for learning values of the network parameters which allow for an approximation of the response of handcrafted detectors, showing that the proposed architecture is able to obtain results comparable with the state of the art. The capability of emulating different detectors allows to deploy a variety of algorithms to dedicated hardware by simply retraining the network. A sensor-based FPGA implementation of the introduced CNN architecture is presented, allowing latency smaller than 1[ms].
2018
The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops
682
690
Paolo Di Febbo, C.D.M. (2018). KCNN: Extremely-Efficient Hardware Keypoint Detection With a Compact Convolutional Neural Network. IEEE/CVF.
Paolo Di Febbo, Carlo Dal Mutto, Kinh Tieu, Stefano Mattoccia
File in questo prodotto:
Eventuali allegati, non sono esposti

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11585/653852
 Attenzione

Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo

Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 14
  • ???jsp.display-item.citation.isi??? 8
social impact