Abstract | ||
---|---|---|
Material recognition from a single image is a challenging problem in the computer vision field due to the lack of reliable and discriminative features. Previous approaches employ off-the-shelf features such as SIFT descriptors or filter bank response to build material recognition systems. The recent success of deep convolutional neural networks (DCNNs) in object recognition motivated us to evaluate their performance in material recognition tasks. In this paper, we tested the generality of several CNN architectures, including VGGNet [31], GoogLeNet [32], Inception V3 [33] and ResNet [10], on two commonly used material datasets: Flickr Material Database (FMD) and Materials IN Context database (MINC). The results show that the best performing CNN architecture, i.e., Inception V3, achieves at least 5% boost on FMD compared with the other networks and almost reaches human's performance. The results on MINC-2500 also exhibit the state-of-the-art level. |
Year | DOI | Venue |
---|---|---|
2017 | 10.1109/UIC-ATC.2017.8397467 | 2017 IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computed, Scalable Computing & Communications, Cloud & Big Data Computing, Internet of People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI) |
Keywords | Field | DocType |
material recognition,deep learning,CNN architectures | Scale-invariant feature transform,Pattern recognition,Convolutional neural network,Convolution,Computer science,MINC,Filter bank,Feature extraction,Artificial intelligence,Discriminative model,Distributed computing,Cognitive neuroscience of visual object recognition | Conference |
ISBN | Citations | PageRank |
978-1-5386-1591-1 | 0 | 0.34 |
References | Authors | |
15 | 5 |
Name | Order | Citations | PageRank |
---|---|---|---|
Xiaowei Shang | 1 | 0 | 0.68 |
ying xu | 2 | 74 | 27.10 |
Lin Qi | 3 | 27 | 8.68 |
Amanuel Hirpa Madessa | 4 | 0 | 0.68 |
Junyu Dong | 5 | 99 | 23.43 |