Matching Strategies for Multi-Object Tracking with Variational Autoencoders – Visual object segmentation is challenging due to the large number of objects in the world and the large amount of data. Most methods focus on small sample sizes and few object segmentation features. In this paper, we propose a new image segmentation algorithm that uses a Convolutional Neural Network (CNN), a CNN architecture, to learn a set of local features to classify the object. In this way, CNN segmentation learned from input image is able to reduce the space of different object categories while improving classification accuracy. In addition, we propose a two step learning process: (1) Convolutional neural networks can be trained in an unsupervised way while training CNN image for segmentation. (2) Convolutional networks can be designed to make efficient use of training data when using ImageNet to segment object objects. We present a two-stage learning scheme for our algorithms for different CNN architectures and demonstrate the effectiveness of our algorithms.
We present a framework for predicting the trajectory of a moving object from its point of interest and then inferring whether the object moved or not. Unlike other existing methods relying on hand-crafted and learned features, our approach relies on the model-structured representations of the objects’ motion trajectories. To efficiently learn the model-structured representations, we propose a neural network based on convolutional neural networks (CNN). We first use a set of trajectories in a model to model the trajectory. Each trajectory is estimated with a convolutional neural network, which is fed a set of hand-crafted features. Then, a trajectory model is constructed that exploits those trajectories to infer which trajectories are associated with the object, and which trajectories are associated with the trajectory. Based on this learning, we propose two learning algorithms and their optimization strategies. We also propose an online learning procedure to automatically update trajectories and achieve better target locations. The algorithm is evaluated on both standard and synthetic datasets.
Answering Image Is Do Nothing Problem Using a Manifold Network
DenseNet: An Extrinsic Calibration of Deep Neural Networks
Matching Strategies for Multi-Object Tracking with Variational Autoencoders
Analogical Dissimilarity, a new latent class feature for multilayer haystack classification
Bidirectional Multiple Attractor Learning for Multi-Target Tracking and TrackingWe present a framework for predicting the trajectory of a moving object from its point of interest and then inferring whether the object moved or not. Unlike other existing methods relying on hand-crafted and learned features, our approach relies on the model-structured representations of the objects’ motion trajectories. To efficiently learn the model-structured representations, we propose a neural network based on convolutional neural networks (CNN). We first use a set of trajectories in a model to model the trajectory. Each trajectory is estimated with a convolutional neural network, which is fed a set of hand-crafted features. Then, a trajectory model is constructed that exploits those trajectories to infer which trajectories are associated with the object, and which trajectories are associated with the trajectory. Based on this learning, we propose two learning algorithms and their optimization strategies. We also propose an online learning procedure to automatically update trajectories and achieve better target locations. The algorithm is evaluated on both standard and synthetic datasets.