Scene recognition is a crucial task in many image processing and computer vision applications - from search by image or automatic image annotation to self-driving cars, drones and surveillance systems. State-of-the-art machine learning algorithms can solve such tasks. Yet, TensorFlow framework makes implementation of these algorithms much easier and faster.
The goal was to build a graph of objects and relationships from the pixels of the given image. The task envisaged locating objects, recognizing them, finding and classifying relationships between the objects.
We have trained an end-to-end model, which can build an object-relationship graph from the image pixel data, without using any intermediate models.
End-to-end learning is a state-of-the-art approach for image analysis and other machine learning tasks. Instead of stacking several models learned independently of each other, end-to-end learning provides a faster way to do all the optimizations by a single training run. Our model had several outputs: objects' positions, their classes, relationships between objects, and types of relationships. We trained and tested the model on the VisualGenome dataset.
PCA
K-means
Decision trees
Linear models
PageRank
Digital filters
DTW
Deep learning
Probabilistic graphical models
CART
ensembles
unsupervised sound segmentation
recurrent models
bayesian approach
probabilistic programming
hmm
alexnet
vgg
vae
PCA
TF-IDF
LDA
SVM
Naive bayes
word2vec
attention models
Hi, we are Sciforce - a company where the integration of various branches of science builds up a powerful force to create robust software solutions. Working at the intersection of Computer Science with other technical, natural and humanitarian sciences let us go beyond traditional IT services and become both technical and scientific forces to our customers.