Object recognition by components and relations between them

The paper’s goal is to develop a methodology and algorithm for the recognition of objects in the environment, keeping the quality with an increasing number of objects. For this purpose, the following problems were solved: recognition of the shape features, estimation of relations between features, and matching between the found features and relations and the defined templates (descriptions of complex and simple objects of the real world). A convolutional neural network is used for the shape feature recognition. In order to train it we used artificially generated images with shape features (3D primitive objects) that were randomly placed on the scene with different properties of their surfaces. The set of relations necessary to recognize objects, which can be represented as a combination of shape features, is formed. Testing on photos of real-world objects showed the ability to recognize real-world objects regardless of their type (in cases where different models and modifications are possible). This paper considers an example of outdoor luminaire recognition. The example shows the algorithm's ability not only to detect an object in the image but also to estimate the position of its components. This solution makes it possible to use the algorithm in the task of object manipulation performed by robotic systems.