 |
Extensive test on large datasets of the SIODISM methodIn this work we are developing a Matlab tool to prove that the SIODISM method is more robust w.r.t. the standard ISM approach. The aim is to demonstrate that the scale information of image descriptors is subject to noise and that using this information to estimate object's scale leads to worst results. |
   |
SIODISM - Scale-Independent Object Detection with an Implicit Shape Model, A. Furlan, D. Marzorati and D. G. Sorrenti, ICDP2009In this paper we propose an improvement to the Implicit Shape Model (ISM) based robust object detection system proposed by Leibe et al. Object detection with ISM allows to approach the classification and tracking in a probabilistic way with multiple hypotheses. Unlike the original approach, our method is independent from object scale in the training sets, and this allows to work with a much smaller training sets and also to avoid to supply information about scale to the trainer. This is done while maintaining the robustness of the original approach. Leibe et al. mentioned a potential solution to overcome the scale problem in the training set, i.e., the usage of the scale produced by the local descriptor. Our proposal is different and developed, up to the preliminary results presented in the paper; moreover, we believe the scale measure from local descriptors mentioned by Leibe et al. to be more noisy.Go to the project's page |
    |
Cognitive scene interpretation based on neural STDP-learning - Master Thesis of Francesco VisinAdvisor: Domenico G. Sorrenti Co-advisor: Axel FurlanThis works grounds on the biologically plausible mechanism by Timothee Masquelier and Simon J. Thorpe based on Spike timing dependent plasticity (STDP), a learning rule that modifies synaptic strength as a function of the relative timing of pre- and postsynaptic spikes. When a neuron is repeatedly presented with similar inputs, STDP concentrates high synaptic weights on afferents that systematically fire early (i.e.: afferents that statistically transmit relevant informations) while postsynaptic spike latencies decrease. This learning rule is used to train a hierarchical feedforward spiking neural network that mimics the ventral visual pathway. Here we use this learning rule in an asynchronous feedforward spiking neural network that mimics the ventral visual pathway for generating intermediate-level visual representations using an unsupervised learning scheme. These representations can then be used very effectively to perform categorization tasks using natural images.The most interesting parts of my work are a mechanism to dynamically create new models (i.e.: neurons) at run time and an interest evaluator. The first one dynamically creates a new model if no known model is good enough to describe the current input pattern. This allows to speed up the learning process by keeping the initial number of neurons very low and increasing it only when needed and permits also a continuous learning of new categories, since presenting the net with new “concepts” will lead to the creation of new models.Go to the project's page |
IRALAB
Informatics and Robotics for Automation
