Multi-stream Convolutional Networks for Indoor Scene Recognition

Rao Muhammad Anwer, Fahad Shahbaz Khan, Jorma Laaksonen, Nazar Zaki

Research output: Chapter in Book/Report/Conference proceedingConference contribution


Convolutional neural networks (CNNs) have recently achieved outstanding results for various vision tasks, including indoor scene understanding. The de facto practice employed by state-of-the-art indoor scene recognition approaches is to use RGB pixel values as input to CNN models that are trained on large amounts of labeled data (ImageNet or Places). Here, we investigate CNN architectures by augmenting RGB images with estimated depth and texture information, as multiple streams, for monocular indoor scene recognition. First, we exploit the recent advancements in the field of depth estimation from monocular images and use the estimated depth information to train a CNN model for learning deep depth features. Second, we train a CNN model to exploit the successful Local Binary Patterns (LBP) by using mapped coded images with explicit LBP encoding to capture texture information available in indoor scenes. We further investigate different fusion strategies to combine the learned deep depth and texture streams with the traditional RGB stream. Comprehensive experiments are performed on three indoor scene classification benchmarks: MIT-67, OCIS and SUN-397. The proposed multi-stream network significantly outperforms the standard RGB network by achieving an absolute gain of 9.3%, 4.7%, 7.3% on the MIT-67, OCIS and SUN-397 datasets respectively.

Original languageEnglish
Title of host publicationComputer Analysis of Images and Patterns - 18th International Conference, CAIP 2019, Proceedings
EditorsMario Vento, Gennaro Percannella
PublisherSpringer Verlag
Number of pages13
ISBN (Print)9783030298876
Publication statusPublished - 2019
Event18th International Conference on Computer Analysis of Images and Patterns, CAIP 2019 - Salerno, Italy
Duration: Sept 3 2019Sept 5 2019

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11678 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349


Conference18th International Conference on Computer Analysis of Images and Patterns, CAIP 2019


  • Depth features
  • Scene recognition
  • Texture features

ASJC Scopus subject areas

  • Theoretical Computer Science
  • General Computer Science


Dive into the research topics of 'Multi-stream Convolutional Networks for Indoor Scene Recognition'. Together they form a unique fingerprint.

Cite this