CoEfNet Project Overview
Computer vision has progressed in many of its fields with the adoption of deep neural networks. Automatic image classification, segmentation and captioning, as well as object recognition and human pose estimation have been especially suitable application areas for deep neural networks. Besides academia, also the industry has expressed considerable interest towards neural network-based solutions, which can be seen in the form of company acquisitions and research investments by car manufacturers. The large-scale deployment of deep neural network -based solutions has however been hindered by computational challenges, which limit their portability to mobile platforms and vehicles. To this extent, the aim of this project is to develop new methodologies for constructing more efficient deep neural networks. The challenge is addressed from two aspects: first, by optimizing existing deep neural networks, and second, by developing efficient neural network architectures from scratch. The partners of the collaborative effort are Aalto University, University of Vaasa, and Tampere University, as well as international collaborators from Grenoble, Oxford and University of Maryland. The project is funded by the Academy of Finland for the duration of 9/2017-2/2022.
Publication Spotlight
LightSAL: Lightweight Sign Agnostic Learning for Implicit Surface Representation
Figure: Point-cloud reconstruction results for 500 training epochs. Left: baseline SAL architecture, Right: proposed LightSAL architecture.
Recently, several works have addressed modeling of 3D shapes using deep neural networks to learn implicit surface representations. Up to now, the majority of works have concentrated on reconstruction quality, paying little or no attention to model size or training time. This work proposes LightSAL, a novel deep convolutional architecture for learning 3D shapes; the proposed work concentrates on efficiency both in network training time and resulting model size. We build on the recent concept of Sign Agnostic Learning for training the proposed network, relying on signed distance fields, with unsigned distance as ground truth. In the experimental section of the paper, we demonstrate that the proposed architecture outperforms previous work in model size and number of required training iterations, while achieving equivalent accuracy. Experiments are based on the D-Faust dataset that contains 41k 3D scans of human shapes. The proposed LightSAL neural network architecture has 75% less trainable parameters, and reduces training time by a factor of 6x compared to the baseline. Moreover, our results show that LightSAL is less prone to overfitting than baseline SAL. Read the full article at arXiv.
Project Consortium
The CoEfNet project is an Academy of Finland consortium project of
Aalto University (Prof. Juho Kannala, juho.kannala@aalto.fi; Google Scholar page, consortium lead)
University of Vaasa (Prof. Jani Boutellier, jani.boutellier@univaasa.fi; Google Scholar page).
Publications
Pedram Ghazi, Antti P. Happonen, Jani Boutellier, Heikki Huttunen (2018)
Embedded Implementation of a Deep Learning Smile Detector
European Workshop on Visual Information Processing (EUVIP)
Bishwo Adhikari, Jukka Peltomäki, Jussi Puura, Heikki Huttunen (2018)
Faster Bounding Box Annotation for Object Detection in Indoor Scenes
European Workshop on Visual Information Processing (EUVIP)
Shayan Gharib, Honain Derrar, Daisuke Niizumi, Tuukka Senttula, Janne Tommola, Toni Heittola, Tuomas Virtanen, Heikki Huttunen (2018)
Acoustic Scene Classification: A Competition Review
IEEE International Workshop on Machine Learning for Signal Processing (MLSP)
Yue Bai, Shuvra S. Bhattacharyya, Antti P. Happonen, Heikki Huttunen (2018)
Elastic Neural Networks: A Scalable Framework for Embedded Computer Vision
European Signal Processing Conference (EUSIPCO)
Yi Zhou, Yue Bai, Shuvra S. Bhattacharyya, Heikki Huttunen (2019)
Elastic Neural Networks for Classification
IEEE Workshop on Artificial Intelligence Circuits and Systems (AICAS), Taipei, Taiwan.
Mir Khan, Heikki Huttunen, Jani Boutellier (2018)
Binarized Convolutional Neural Networks for Efficient Inference on GPUs
European Signal Processing Conference (EUSIPCO)
Ke Chen, Kui Jia, Heikki Huttunen, Jiri Matas, Joni Kämäräinen (2019)
Cumulative Attribute Space Regression for Head Pose Estimation and Color Constancy
Pattern Recognition. 87, p. 29-37
Payvar S, Khan M, Stahl R, Müller-Gritschneder D, Boutellier J (2019) [Best paper award, 2nd prize]
Neural Network based Vehicle Image Classification for IoT Devices
IEEE Workshop on Signal Processing Systems (SiPS)
Khan M, Lunnikivi H, Huttunen H, Boutellier J (2019) [Best student paper candidate]
Comparing Optimization Methods of Neural Networks for Real-time Inference
European Signal Processing Conference (EUSIPCO)
Ferranti L (2019)
Confidence estimation in image based localization
Master’s thesis, Tampere University
Ma Y, Wu J, Bhattacharyya S, Boutellier J (2020)
Decidable Variable-Rate Dataflow for Heterogeneous Signal Processing Systems
International Conference on Acoustics, Speech, and Signal Processing (ICASSP).
Ferranti L, Li X, Boutellier J, Kannala J (2020)
Can You Trust Your Pose? Confidence Estimation in Visual Localization
International Conference on Pattern Recognition (ICPR).
Basher A, Sarmad M and Boutellier J (2021)
LightSAL: Lightweight Sign Agnostic Learning for Implicit Surface Representation
arXiv preprint arXiv:2103.14273.
Basher A and Boutellier J (2022)
Convolutional Neural Network-based Efficient Dense Point Cloud Generation using Unsigned Distance Fields
arXiv preprint arXiv:2203.11537.
Boutellier J, Ma Y, Wu J, Khan M and Bhattacharyya SS (2022)
VR-PRUNE: Decidable Variable-Rate Dataflow for Signal Processing Systems
IEEE Transactions on Signal Processing
More CoEfNet papers can be found on the webpage of Juho Kannala (Aalto University).
Presentations
Luca Ferranti
INTO indoor localization seminar
Nov 29, 2019
Yujunrong Ma
Decidable Variable-Rate Dataflow for Heterogeneous Signal Processing Systems
Video at ICASSP 2020
Luca Ferranti
DELTA autumn workshop
Nov 12, 2020
Luca Ferranti
FCAI AI day
Nov 26, 2020
Luca Ferranti
International Conference on Pattern Recognition
Jan 15, 2021
Abol Basher
CSC Lumi roadshow
Feb 19, 2021
Abol Basher
DELTA spring workshop
June 8-9, 2021
Abol Basher
FCAI SIG CV webinar
August 25, 2021