Highlight

Social information filtering: algorithms for automating “word of mouth”

Introduces social information filtering, making recommendations from similarities between users' interest profiles, tested with the Ringo system.

This paper describes social information filtering, a technique for personalized recommendations from any database based on similarities between a user's interest profile and those of other users. The authors implemented Ringo, a networked system that recommends music albums and artists, with a database of users and artists that grows dynamically as more people use it. Four different recommendation algorithms were tested and compared, and the paper presents quantitative and qualitative results from use of Ringo by more than 2000 people.

Based on: Social information filtering: algorithms for automating “word of mouth” · International Conference on Human Factors in Computing Systems

Curated by Aramai Editorial

Read summary →
Highlight

Ward’s Hierarchical Agglomerative Clustering Method: Which Algorithms Implement Ward’s Criterion?

Clarifies that two commonly used algorithms both claiming to implement Ward's clustering criterion actually give different results.

Ward's error sum of squares hierarchical clustering method, first described in 1963, has been very widely used and generalized. Two algorithms found in the literature and software both claim to implement Ward's method, yet when applied to the same distance matrix they produce different results, because only one preserves Ward's criterion. Through survey work and case studies, the paper clarifies this distinction to aid anyone developing data analysis software using Ward's hierarchical clustering method.

Based on: Ward’s Hierarchical Agglomerative Clustering Method: Which Algorithms Implement Ward’s Criterion? · Journal of Classification

Curated by Aramai Editorial

Read summary →
Highlight

Occupancy Networks: Learning 3D Reconstruction in Function Space

Introduces Occupancy Networks, representing 3D surfaces implicitly as the continuous decision boundary of a neural classifier.

Learning-based 3D reconstruction lacks a representation that is memory-efficient yet captures high-resolution geometry of arbitrary topology, so many methods produce coarse geometry or work only in restricted domains. Occupancy Networks implicitly represent the 3D surface as the continuous decision boundary of a deep neural network classifier, encoding output at effectively infinite resolution without excessive memory. Experiments show competitive results for 3D reconstruction from single images, noisy point clouds, and coarse voxel grids.

Based on: Occupancy Networks: Learning 3D Reconstruction in Function Space · Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

Progressive meshes

Introduces the progressive mesh, a lossless continuous-resolution representation for storing, transmitting, and simplifying triangle meshes.

Highly detailed triangle meshes strain rendering, transmission, and storage. This paper introduces the progressive mesh (PM) representation, an efficient, lossless, continuous-resolution scheme for arbitrary triangle meshes that supports smooth level-of-detail geomorphing, progressive transmission, mesh compression, and selective refinement. It also presents a mesh simplification procedure for building a PM from an arbitrary mesh, optimizing to preserve not just geometry but overall appearance defined by attributes like material identifiers, colors, normals, and texture coordinates.

Based on: Progressive meshes · International Conference on Computer Graphics and Interactive Techniques

Curated by Aramai Editorial

Read summary →
Highlight

CenterNet: Keypoint Triplets for Object Detection

Proposes CenterNet, detecting each object as a keypoint triplet with center and cascade corner pooling to cut false bounding boxes.

Keypoint-based detectors often produce many incorrect bounding boxes because they lack assessment inside cropped regions. Building on the one-stage detector CornerNet, CenterNet detects each object as a triplet rather than a pair of keypoints, improving precision and recall. Two custom modules, cascade corner pooling and center pooling, enrich corner information and add recognizable central-region cues. On MS-COCO, CenterNet reaches 47.0% AP, outperforming all existing one-stage detectors by at least 4.9%, with faster inference than top two-stage detectors and comparable performance.

Based on: CenterNet: Keypoint Triplets for Object Detection · IEEE International Conference on Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

MnasNet: Platform-Aware Neural Architecture Search for Mobile

Introduces MNAS, an automated mobile neural architecture search that folds real measured on-device latency into the search objective.

Designing accurate yet small and fast mobile CNNs is hard given many architectural trade-offs. MnasNet proposes an automated mobile neural architecture search that incorporates model latency directly into the objective, measuring real inference latency by running models on phones rather than using FLOPS proxies. A factorized hierarchical search space encourages layer diversity. On ImageNet, MnasNet reaches 75.2% top-1 accuracy at 78ms on a Pixel phone, 1.8x faster than MobileNetV2 with higher accuracy, and also improves COCO detection mAP over MobileNets.

Based on: MnasNet: Platform-Aware Neural Architecture Search for Mobile · Computer Vision and Pattern Recognition

Curated by Aramai Editorial

Read summary →
Highlight

Accelerating the Super-Resolution Convolutional Neural Network

Proposes a compact hourglass-shaped CNN (FSRCNN) that accelerates SRCNN over 40x while improving super-resolution restoration quality.

SRCNN achieved strong image super-resolution results but its high computational cost prevents real-time use. This paper accelerates it with a compact hourglass-shaped CNN using three redesigns: a deconvolution layer at the end so mapping is learned directly from the original low-resolution image, a reformulated mapping layer that shrinks then expands feature dimensions, and smaller filters with more mapping layers. The model runs over 40x faster with superior restoration quality, achieves real-time performance on a generic CPU, and supports a transfer strategy across upscaling factors.

Based on: Accelerating the Super-Resolution Convolutional Neural Network · European Conference on Computer Vision

Curated by Aramai Editorial

Read summary →
Highlight

Handbook of Markov Chain Monte Carlo

An edited handbook collecting foundational and applied chapters on Markov chain Monte Carlo methods across statistics and science.

This handbook is a collected volume of chapters covering Markov chain Monte Carlo (MCMC) methods, spanning introductions, history, and advanced techniques. Topics include reversible jump MCMC, adaptive MCMC and optimal proposal distributions, Hamiltonian dynamics, convergence monitoring, exact sampling, importance sampling and tempering, and likelihood-free MCMC. It also presents diverse applications including genetics, functional MRI, astrophysics, ecology, spatial data, state space models, educational research, fisheries, and epidemiology.

Based on: Handbook of Markov Chain Monte Carlo · arXiv

Curated by Aramai Editorial

Read summary →
Highlight

A survey of transfer learning

A survey formally defining transfer learning and reviewing current solutions and applications for learning across differing domains.

Traditional machine learning assumes training and testing data share the same domain, feature space, and distribution, but this often fails in real-world settings where data is expensive or hard to collect. Transfer learning addresses this by building high-performance learners using more easily obtained data from different domains. This survey formally defines transfer learning, reviews current solutions and applications, and lists software downloads. The surveyed solutions are independent of data size and applicable to big data environments.

Based on: A survey of transfer learning · Journal of Big Data

Curated by Aramai Editorial

Read summary →
Highlight

Object Detection in 20 Years: A Survey

A comprehensive survey reviewing over 25 years of object detection technical evolution, from 1990s methods to deep-learning-era detectors.

This article surveys object detection, one of computer vision's most fundamental and challenging problems, tracing its rapid technological evolution across more than a quarter-century (1990s to 2022). It frames today's detection as a deep-learning-driven revolution while acknowledging the ingenious early designs. Coverage includes milestone detectors, detection datasets, evaluation metrics, fundamental building blocks of detection systems, speedup techniques, and recent state-of-the-art methods.

Based on: Object Detection in 20 Years: A Survey · Proceedings of the IEEE

Curated by Aramai Editorial

Read summary →
Highlight

UNet++: Redesigning Skip Connections to Exploit Multiscale Features in Image Segmentation

Proposes UNet++, redesigning skip connections and using nested U-Nets to improve semantic and instance image segmentation.

State-of-the-art medical image segmentation models are U-Net and FCN variants, but their optimal depth is unknown and their skip connections restrictively fuse only same-scale encoder-decoder features. UNet++ adds an efficient ensemble of U-Nets of varying depths that share an encoder and co-learn via deep supervision, redesigned skip connections aggregating features across semantic scales, and a pruning scheme to speed inference. Across six CT, MRI, and electron microscopy datasets, UNet++ outperforms baselines and, as Mask RCNN++, surpasses Mask R-CNN for instance segmentation.

Based on: UNet++: Redesigning Skip Connections to Exploit Multiscale Features in Image Segmentation · IEEE Transactions on Medical Imaging

Curated by Aramai Editorial

Read summary →
Highlight

Objects as Points

Introduces CenterNet, which models objects as center keypoints and regresses their properties for fast, accurate detection.

Most successful object detectors enumerate a nearly exhaustive list of candidate object locations and classify each, which is wasteful and needs post-processing. This paper instead models an object as a single point, the center of its bounding box, using keypoint estimation to find centers and regressing properties like size, 3D location, orientation, and pose. CenterNet is end-to-end differentiable, simpler, faster, and more accurate than box-based methods. On MS COCO it reaches 28.1% AP at 142 FPS and 45.1% AP with multi-scale testing, plus KITTI 3D and COCO pose.

Based on: Objects as Points · arXiv.org

Curated by Aramai Editorial

Read summary →