In this paper we describe development of a computer vision system for accurate detection and classification of clothes for e-commerce images. We present a set of experiments on well established architectures of convolutional neural networks, including Residual networks, SqueezeNet and Single Shot MultiBox Detector (SSD). The clothes detection network was trained and tested on DeepFashion dataset, which contains box annotations for locations of clothes. Classification task was evaluated on a set of images of dresses that were collected from online shops. Ground truth labels were inferred from shop items metadata for five different attributes, including color, pattern, sleeve, neckline and hemline, each consisting of several possible classes. Automatic gathering of labels resulted in an average of 83% rate of correct labels. In the experiments we evaluate the impact on classification accuracy of a set of potential improvements, including data augmentation by generating diverse backgrounds, increasing the size of the network and using ensembles. We analyse the accuracy improvements with respect to the processing efficiency. Finally, we present the achieved accuracy rates in the clothes detection task and outline the most successful network configurations for dresses classification.

In this paper, we propose a distributed system for point cloud processing and transferring them via computer network regarding to effectiveness-related requirements. We discuss the comparison of point cloud filters focusing on their usage for streaming optimization. For the filtering step of the stream pipeline processing we evaluate four filters: Voxel Grid, Radial Outliner Remover, Statistical Outlier Removal and Pass Through. For each of the filters we perform a series of tests for evaluating the impact on the point cloud size and transmitting frequency (analysed for various fps ratio). We present results of the optimization process used for point cloud consolidation in a distributed environment. We describe the processing of the point clouds before and after the transmission. Pre- and post-processing allow the user to send the cloud via network without any delays. The proposed pre-processing compression of the cloud and the post-processing reconstruction of it are focused on assuring that the end-user application obtains the cloud with a given precision.

The presented paper investigates the problem of endoscopic bleeding detection in endoscopic videos in the form of a binary image classification task. A set of definitions of high-level visual features of endoscopic bleeding is introduced, which incorporates domain knowledge from the field. The high-level features are coupled with respective feature descriptors, enabling automatic capture of the features using image processing methods. Each of the proposed feature descriptors outputs a feature activation map in the form of a grayscale image. Acquired feature maps can be appended in a straightforward way to the original color channels of the input image and passed to the input of a convolutional neural network during the training and inference steps. An experimental evaluation is conducted to compare the classification ROC AUC of feature-extended convolutional neural network models with baseline models using regular color image inputs. The advantage of feature-extended models is demonstrated for the Resnet and VGG convolutional neural network architectures.

Aims: To define a set of high-level visual features of endoscopic bleeding and evaluate their capabilities for potential use in automatic bleeding detection. Study Design: Experimental study. Place and Duration of Study: Department of Computer Architecture, Faculty of Electronics, Telecommunications and Informatics, Gdansk University of Technology, between March 2014 and May 2014. Methodology: The features have been established and evaluated on images extracted from 135 different endoscopic examinations. A set of measures were proposed in order to evaluate the quality of the features. The potential contribution of the features for bleeding detection is evaluated by utilizing an SVM classifier on the vectors constructed by manually assigning the features values to the endoscopic images. Potential limitations of the features' implementations are simulated by introducing noise to the assigned values. Finally, the recognition efficiency of the method is compared to implementations of two bleeding detection algorithms from the literature. Results: The experiments show a potential of bleeding detection efficiency of 95%, assuming perfect features' implementations, and 90% when features' evaluation efficiency of 95% is assumed. Conclusion: By introducing the visual features, the challenging problem of detecting bleeding in endoscopic images is split into a set of smaller image processing problems. The divide-and-conquer approach based on the proposed visual features is capable of achieving high detection efficiency; at the same time providing additional benefits of a more apparent processing scheme, and a possibility of justifying the decisions in a human-readable form.