Cardiff University | Prifysgol Caerdydd ORCA
Online Research @ Cardiff 
WelshClear Cookie - decide language by browser settings

Deriving visual semantics from spatial context: an adaptation of LSA and Word2Vec to generate object and scene embeddings from images

Treder, Matthias, Mayor-Torres, Juan and Teufel, Christoph 2020. Deriving visual semantics from spatial context: an adaptation of LSA and Word2Vec to generate object and scene embeddings from images. [Online]. Cornell University. Available at: https://arxiv.org/abs/2009.09384

[thumbnail of 2009.09384.pdf]
Preview
PDF - Submitted Pre-Print Version
Download (5MB) | Preview

Abstract

Embeddings are an important tool for the representation of word meaning. Their effectiveness rests on the distributional hypothesis: words that occur in the same context carry similar semantic information. Here, we adapt this approach to index visual semantics in images of scenes. To this end, we formulate a distributional hypothesis for objects and scenes: Scenes that contain the same objects (object context) are semantically related. Similarly, objects that appear in the same spatial context (within a scene or subregions of a scene) are semantically related. We develop two approaches for learning object and scene embeddings from annotated images. In the first approach, we adapt LSA and Word2vec's Skipgram and CBOW models to generate two sets of embeddings from object co-occurrences in whole images, one for objects and one for scenes. The representational space spanned by these embeddings suggests that the distributional hypothesis holds for images. In an initial application of this approach, we show that our image-based embeddings improve scene classification models such as ResNet18 and VGG-11 (3.72\% improvement on Top5 accuracy, 4.56\% improvement on Top1 accuracy). In the second approach, rather than analyzing whole images of scenes, we focus on co-occurrences of objects within subregions of an image. We illustrate that this method yields a sensible hierarchical decomposition of a scene into collections of semantically related objects. Overall, these results suggest that object and scene embeddings from object co-occurrences and spatial context yield semantically meaningful representations as well as computational improvements for downstream applications such as scene classification.

Item Type: Website Content
Date Type: Submission
Status: Submitted
Schools: Psychology
Cardiff University Brain Research Imaging Centre (CUBRIC)
Publisher: Cornell University
Date of First Compliant Deposit: 22 September 2020
Date of Acceptance: 22 September 2020
Last Modified: 28 Apr 2022 11:21
URI: https://orca.cardiff.ac.uk/id/eprint/135003

Actions (repository staff only)

Edit Item Edit Item

Downloads

Downloads per month over past year

View more statistics