Abstract: A new deep visual odometry (VO) method that selects memory and refines poses considers global information. Existing learning-based methods treat VO as a pure tracking problem by recovering camera poses from image snippets, resulting…
Abstract: Social network analyses and applications face the user alignment problem, which matches users across networks. State-of-the-art methods embed users into the low-dimensional representation space, where their features are preserved, and establish user correspondence based…
Abstract: IoT is transforming human life. However, the rapid adoption of IoT makes cyberspace more vulnerable, especially to IoT-based attacks on cyber-physical systems. With billions of IoT devices, detecting and preventing these attacks is crucial….
Abstract: The distance metric that measures object dissimilarity is crucial to categorical data clustering. Most clustering methods ignore the relative order of ordinal values when calculating dissimilarity between nominal and ordinal attributes. Interdependence between nominal…
Abstract: Low-rank minimization recovers a minimum-rank matrix under linear system constraint. Recommender systems, signal processing, and video denoising use it. Nuclear norm minimization dominates. However, it ignores target matrix singular values. Nonconvex low-rank regularizers solve…
Abstract: Multi-source domain adaptation requires reducing domain discrepancy between source domains and target domains and assessing domain relevance to determine how much knowledge should be transferred. Most previous approaches ignored domain discrepancies and relevance. This…
Abstract: Training a deep neural network without overfitting is difficult with limited training data. This article proposes an adversarial data augmentation network (ADAN) based on GANs to solve this problem. GAN, autoencoder, and auxiliary classifier…
Abstract: Hypergraph learning teaches hypergraph structures. Due to its flexibility and ability to model complex data correlation, hypergraph learning has gained popularity. First, we review distance-based, representation-based, attribute-based, and network-based hypergraph generation literature. Transductive, inductive,…
Abstract: Deep learning detectors efficiently detect multi-scale objects in one image. Recent works like FPN and SSD use feature maps from multiple layers with different spatial resolutions to detect objects at different scales, e.g., high-resolution…
Abstract: Feature selection eliminates unimportant features. Embedded feature selection methods, which learn feature weights during classifier training, have garnered attention recently. Traditional embedded methods only optimize all features combinatorially. They sometimes select weakly relevant features…
Abstract: Real-world image classification problems often have class imbalance, with some classes having abundant data and others not. In this case, classifier representations are biased toward the majority classes and learning proper features is difficult,…
Abstract: Multi-modal hashing, a promising method for scaling cross-view retrieval, is gaining attention as multimedia data grows exponentially. Most multi-modal hashing methods either unnaturally split the learning process into two stages or treat the discrete…
WhatsApp us