Search

Recently, a common starting point for solving complex unsupervised image classification tasks is to use generic features, extracted with deep Convolutional Neural Networks (CNN) pretrained on a large and versatile dataset (ImageNet). However, in most research, the CNN architecture for feature extraction is chosen arbitrarily, without justification. This paper aims at providing insight on the use of pretrained CNN features for image clustering (IC). First, extensive experiments are conducted and show that, for a given dataset, the choice of the CNN architecture for feature extraction has a huge impact on the final clustering. These experiments also demonstrate that proper extractor selection for a given IC task is difficult. To solve this issue, we propose to rephrase the IC problem as a multi-view clustering (MVC) problem that considers features extracted from different architectures as different "views" of the same data. This approach is based on the assumption that information contained in the different CNN may be complementary, even when pretrained on the same data. We then propose a multi-input neural network architecture that is trained end-to-end to solve the MVC problem effectively. This approach is tested on nine natural image datasets, and produces state-of-the-art results for IC., Comment: 21 pages, 16 figures, 10 tables, preprint of our paper published in Neurocomputing

Object Detection (OD) is an important task in Computer Vision with many practical applications. For some use cases, OD must be done on videos, where the object of interest has a periodic motion. In this paper, we formalize the problem of periodic OD, which consists in improving the performance of an OD model in the specific case where the object of interest is repeating similar spatio-temporal trajectories with respect to the video frames. The proposed approach is based on training a Gaussian Process to model the periodic motion, and use it to filter out the erroneous predictions of the OD model. By simulating various OD models and periodic trajectories, we demonstrate that this filtering approach, which is entirely data-driven, improves the detection performance by a large margin., 8 pages, 11 figures, 1 table Accepted as a full paper at ICMLA 2020 (19th IEEE International Conference On Machine Learning And Applications)

With the major adoption of automation for cities security, person re-identification (Re-ID) has been extensively studied recently. In this paper, we argue that the current way of studying person re-identification, i.e. by trying to re-identify a person within already detected and pre-cropped images of people, is not sufficient to implement practical security applications, where the inputs to the system are the full frames of the video streams. To support this claim, we introduce the Full Frame Person Re-ID setting (FF-PRID) and define specific metrics to evaluate FF-PRID implementations. To improve robustness, we also formalize the hybrid human-machine collaboration framework, which is inherent to any Re-ID security applications. To demonstrate the importance of considering the FF-PRID setting, we build an experiment showing that combining a good people detection network with a good Re-ID model does not necessarily produce good results for the final application. This underlines a failure of the current formulation in assessing the quality of a Re-ID model and justifies the use of different metrics. We hope that this work will motivate the research community to consider the full problem in order to develop algorithms that are better suited to real-world scenarios., 7 pages, 9 figures, This paper is under consideration at Pattern Recognition Letters

The contribution of this paper is twofold. First, a new data driven approach for predicting the Covid-19 pandemic dynamics is introduced. The second contribution consists in reporting and discussing the results that were obtained with this approach for the Brazilian states, with predictions starting as of 4 May 2020. As a preliminary study, we first used an Long Short Term Memory for Data Training-SAE (LSTM-SAE) network model. Although this first approach led to somewhat disappointing results, it served as a good baseline for testing other ANN types. Subsequently, in order to identify relevant countries and regions to be used for training ANN models, we conduct a clustering of the world&rsquo, s regions where the pandemic is at an advanced stage. This clustering is based on manually engineered features representing a country&rsquo, s response to the early spread of the pandemic, and the different clusters obtained are used to select the relevant countries for training the models. The final models retained are Modified Auto-Encoder networks, that are trained on these clusters and learn to predict future data for Brazilian states. These predictions are used to estimate important statistics about the disease, such as peaks and number of confirmed cases. Finally, curve fitting is carried out to find the distribution that best fits the outputs of the MAE, and to refine the estimates of the peaks of the pandemic. Predicted numbers reach a total of more than one million infected Brazilians, distributed among the different states, with S&atilde, o Paulo leading with about 150 thousand confirmed cases predicted. The results indicate that the pandemic is still growing in Brazil, with most states peaks of infection estimated in the second half of May 2020. The estimated end of the pandemics (97% of cases reaching an outcome) spread between June and the end of August 2020, depending on the states.

This paper has a twofold contribution. The first is a data driven approach for predicting the Covid-19 pandemic dynamics, based on data from more advanced countries. The second is to report and discuss the results obtained with this approach for Brazilian states, as of May 4th, 2020. We start by presenting preliminary results obtained by training an LSTM-SAE network, which are somewhat disappointing. Then, our main approach consists in an initial clustering of the world regions for which data is available and where the pandemic is at an advanced stage, based on a set of manually engineered features representing a country’s response to the early spread of the pandemic. A Modified Auto-Encoder network is then trained from these clusters and learns to predict future data for Brazilian states. These predictions are used to estimate important statistics about the disease, such as peaks. Finally, curve fitting is carried out on the predictions in order to find the distribution that best fits the outputs of the MAE, and to refine the estimates of the peaks of the pandemic. Results indicate that the pandemic is still growing in Brazil, with most states peaks of infection estimated between the 25th of April and the 19th of May 2020. Predicted numbers reach a total of 240 thousand infected Brazilians, distributed among the different states, with São Paulo leading with almost 65 thousand estimated, confirmed cases. The estimated end of the pandemics (with 97% of cases reaching an outcome) starts as of May 28th for some states and rests through August 14th, 2020.

Globally, conversion of pristine areas to anthropogenic landscapes is one of the main causes of ecosystem service losses. Land uses associated with urbanization and farming can be major sources of pollution to freshwaters promoting artificial inputs of several elements, leading to impaired water quality. However, how the effects of land use on freshwater quality are contingent on properties of the local landscape and climate is still poorly understood. The aim of this study was to evaluate the effects of landscape properties (morphometric measurements of lakes and their catchments), precipitation patterns, and land use properties (extent and proximity of the land use to freshwaters) on water quality of 98 natural lakes and reservoirs in northeast Brazil. Water quality impairment (WQI) was expressed as a composite variable incorporating parameters correlated with eutrophication including nitrogen (N), phosphorus (P) and Chlorophyll-a concentration. Regression tree analysis showed that WQI is mainly related to highly impacted “buffer areas”. However, the effects of land use in these adjacent lands were contingent on precipitation variability for 13% of waterbodies and on surface area of the buffer in relation to the volume of waterbody (BA:Vol) for 87% of waterbodies. Overall, effects on WQI originating from the land use in the adjacent portion of the lake were amplified by high precipitation variability for ecosystems with highly impacted buffer areas and by high BA:Vol for ecosystems with less impacted buffer areas, indicating that ecosystems subjected to intense episodic rainfall events (e.g. storms) and higher buffer areas relative to aquatic ecosystem size (i.e. small waterbodies) are more susceptible to impacts of land use. Land use at the catchment scale was important for the largest ecosystems. Thus, our findings point toward the need for considering a holistic approach to managing water quality, which includes watershed management within the context of climate change.

Unlike classification, position labels cannot be assigned manually by humans. For this reason, generating supervision for precise object localization is a hard task. This paper details a method to create large datasets for 3D object localization, with real world images, using an industrial robot to generate position labels. By knowledge of the geometry of the robot, we are able to automatically synchronize the images of the two cameras and the object 3D position. We applied it to generate a screw-driver localization dataset with stereo images, using a KUKA LBR iiwa robot. This dataset could then be used to train a CNN regressor to learn end-to-end stereo object localization from a set of two standard uncalibrated cameras.

An understanding of the nature of objects could help robots to solve both high-level abstract tasks and improve performance at lower-level concrete tasks. Although deep learning has facilitated progress in image understanding, a robot's performance in problems like object recognition often depends on the angle from which the object is observed. Traditionally, robot sorting tasks rely on a fixed top-down view of an object. By changing its viewing angle, a robot can select a more semantically informative view leading to better performance for object recognition. In this paper, we introduce the problem of semantic view selection, which seeks to find good camera poses to gain semantic knowledge about an observed object. We propose a conceptual formulation of the problem, together with a solvable relaxation based on clustering. We then present a new image dataset consisting of around 10k images representing various views of 144 objects under different poses. Finally we use this dataset to propose a first solution to the problem by training a neural network to predict a "semantic score" from a top view image and camera pose. The views predicted to have higher scores are then shown to provide better clustering results than fixed top-down views., Comment: 6 pages double columns, 5 figures, 3 tables, Accepted for presentation at IROS 2018, Madrid, Spain (46% acceptance)

This paper aims at providing insight on the transferability of deep CNN features to unsupervised problems. We study the impact of different pretrained CNN feature extractors on the problem of image set clustering for object classification as well as fine-grained classification. We propose a rather straightforward pipeline combining deep-feature extraction using a CNN pretrained on ImageNet and a classic clustering algorithm to classify sets of images. This approach is compared to state-of-the-art algorithms in image-clustering and provides better results. These results strengthen the belief that supervised training of deep CNN on large datasets, with a large variability of classes, extracts better features than most carefully designed engineering approaches, even for unsupervised tasks. We also validate our approach on a robotic application, consisting in sorting and storing objects smartly based on clustering., Comment: 10 pages, 2 figures, 4 tables. Proceedings of AIFU 2018, Melbourne, Australia

This paper describes a method for clustering data that are spread out over large regions and which dimensions are on different scales of measurement. Such an algorithm was developed to implement a robotics application consisting in sorting and storing objects in an unsupervised way. The toy dataset used to validate such application consists of Lego bricks of different shapes and colors. The uncontrolled lighting conditions together with the use of RGB color features, respectively involve data with a large spread and different levels of measurement between data dimensions. To overcome the combination of these two characteristics in the data, we have developed a new weighted K-means algorithm, called gap-ratio K-means, which consists in weighting each dimension of the feature space before running the K-means algorithm. The weight associated with a feature is proportional to the ratio of the biggest gap between two consecutive data points, and the average of all the other gaps. This method is compared with two other variants of K-means on the Lego bricks clustering problem as well as two other common classification datasets., Comment: 13 pages, 6 figures, 2 tables. This paper is under the review process for AIAP 2017

Recent methods of Reinforcement Learning have enabled to solve difficult, high dimensional, robotic tasks under unknown dynamics using iterative Linear Quadratic Gaussian control theory. These algorithms are based on building a local time-varying linear model of the dynamics from data gathered through interaction with the environment. In such tasks, the cost function is often expressed directly in terms of the state and control variables so that it can be locally quadratized to run the algorithm. If the cost is expressed in terms of other variables, a model is required to compute the cost function from the variables manipulated. We propose a method to learn the cost function directly from the data, in the same way as for the dynamics. This way, the cost function can be defined in terms of any measurable quantity and thus can be chosen more appropriately for the task to be carried out. With our method, any sensor information can be used to design the cost function. We demonstrate the efficiency of this method through simulating, with the V-REP software, the learning of a Cartesian positioning task on several industrial robots with different characteristics. The robots are controlled in joint space and no model is provided a priori. Our results are compared with another model free technique, consisting in writing the cost function as a state variable.