Eraldo Ribeiro

Research

My research develops computational methods for learning and extracting structure from complex visual and temporal data. Images, video, biological signals, environmental imagery, and social systems are high-dimensional — yet structured. The goal is to recover that hidden structure and turn it into scientific insight.

Research highlights

Visual summaries of selected results. Click an infographic to enlarge.

Infographic: Compressing Deep ResNets for Real-Time Embedded AI — a two-step pipeline of APoZ-based weight pruning and teacher–student knowledge distillation, achieving a 20× model-size reduction (99.2 MB to 4.9 MB) and 5.5× faster inference on a Jetson Nano.

Compressing Deep ResNets for Real-Time Embedded AI

A two-step pipeline — selective weight pruning via Average Percentage of Zeros (APoZ) followed by teacher–student knowledge distillation with a deep-layer cosine-similarity loss — shrinks a ResNet-50 by 20× (99.2 → 4.9 MB) and runs 5.5× faster on a Jetson Nano, with only a small accuracy trade-off on head-pose estimation.

N. Aghli, E. Ribeiro. “Combining Weight Pruning and Knowledge Distillation for CNN Compression.” IEEE/CVF CVPR Workshops, 2021. · See publications

Infographic: ReScape — unlocking 140 years of coral reef data. A seven-function pipeline (lens calibration and distortion removal, horizon and roll alignment, and brute-force inverse-perspective mapping) rectifies historical oblique reef photos from 1885 onward into top-down scientific data, yielding a ~90% reduction in effort versus traditional 1-square-meter photoquadrats.

ReScape: Unlocking 140 Years of Coral-Reef Data

Millions of perspective-distorted historical reef photographs (from 1885 onward) were unusable for measurement. ReScape applies lens-distortion correction, horizon and roll alignment, and brute-force inverse-perspective mapping via vanishing points to rectify oblique images into calibrated top-down views — cutting effort by ~90% versus traditional 1 m² photoquadrats, and delivered as a free app needing no coding.

Z. Ferris, E. Ribeiro, T. Nagata, R. van Woesik. “ReScape: Transforming Coral-Reefscape Images for Quantitative Analysis.” Scientific Reports, 2024. · Source code

Infographic: Mapping the invisible — a low-cost approach to air-pollution estimation. Official monitoring is expensive and urban-centric, leaving rural areas without data; a road-based regression model using OpenStreetMap road structure predicts pollution across England and Wales from only a tiny fraction of districts.

Low-Cost Air-Pollution Estimation from Road Networks

Official air-quality monitoring is expensive and urban-centric, leaving rural areas without data. Using only open OpenStreetMap road data, a simple linear regression predicts ambient pollution across England and Wales — with the length of “track” and “unclassified” roads acting as a surprisingly strong proxy. A tiny fraction of districts is enough to estimate levels region-wide, offering a low-barrier tool for policymakers, especially in low-income countries.

L. Berrisfeld, E. Ribeiro, R. Menezes. “Estimating Annual Ambient Air Pollution Using Structural Properties of Road Networks.” Environment and Planning B, 2024. · Complex Systems & Network Analysis

Infographic: The travelling waves of crime — uncovering the dynamic rhythm of cities. City-level data suggests a stable yearly crime cycle, but a localized wavelet analysis of twelve cities reveals non-stationary waves that migrate between neighborhoods; the number of active regions stays constant while their locations perpetually shift.

The Travelling Waves of Crime

Aggregated city data suggests a stable yearly crime rhythm — an apparent urban equilibrium. Zooming into smaller regions across twelve cities and applying wavelet analysis reveals the opposite: the seasonal “waves” are non-stationary and constantly migrate between neighborhoods. The number of areas in a cycle stays roughly constant, but which neighborhoods are involved is in perpetual flux — evidence that cities behave as complex systems in disequilibrium, shaped by bottom-up interactions rather than central stability.

M. Oliveira, E. Ribeiro, C. Bastos-Filho, R. Menezes. “Spatio-temporal Variations in the Urban Rhythm: The Travelling Waves of Crime.” EPJ Data Science, 2018. · Complex Systems & Network Analysis

From data to insight

A common thread runs through my work: discover the latent structure in raw data, then reason about it.

  1. Complex dataimages · video · signals · networks
  2. Structure discoverygeometry · motion · interactions · networks
  3. Scientific insightunderstanding in science & society

Main themes

Five connected lines of research.

01

Geometry & Visual Perception

Recovering 3-D structure and surface geometry from a single image — inferring local surface orientation from the way regular textures distort under perspective, estimating pose, and recovering surface normals using spectral (Fourier) methods.

  • shape-from-texture
  • pose estimation
  • spectral methods
Representative publications
  • E. Ribeiro, E. R. Hancock. “Shape from Periodic Texture Using the Eigenvectors of Local Affine Distortion.” IEEE TPAMI, 2002.
  • E. Ribeiro, E. R. Hancock. “Estimating the Perspective Pose of Texture Planes Using Spectral Analysis on the Unit Sphere.” Pattern Recognition, 2002.
  • E. Ribeiro, E. R. Hancock. “Adapting Spectral Scale for Shape from Texture.” ECCV, 2000.
  • E. Ribeiro, F. Sartori, E. R. Hancock. “An Evidence Combining Approach to Shape-from-Shading.” ICPR, 2002.
02

Temporal Understanding of Visual Data

Making sense of motion in video — tracking objects through occlusion, recognizing human actions, and modeling how a person interacts with objects over time. When someone handles an object, their motion is constrained by that object’s shape and affordance — a cue for actor–object interaction recognition.

  • tracking
  • motion
  • action recognition
  • activity discovery
Representative publications
  • R. Filipovych, E. Ribeiro. “Recognizing Primitive Interactions by Exploring Actor-Object States.” IEEE CVPR, 2008.
  • R. Filipovych, E. Ribeiro. “Robust Sequence Alignment for Actor-Object Interaction Recognition: Discovering Actor-Object States.” Computer Vision and Image Understanding, 2011.
  • R. Filipovych, E. Ribeiro. “Learning Human Motion Models from Unsegmented Videos.” IEEE CVPR, 2008.
  • D. Kular, E. Ribeiro. “Analyzing Activities in Videos Using Latent Dirichlet Allocation and Granger Causality.” ISVC, 2015.
  • I. Bogun, E. Ribeiro. “Object-Aware Tracking.” ICPR, 2016.
03

Recognition & Classification in Complex Domains

Bringing recognition methods to challenging scientific data — identifying pollen grains in optical microscopy, classifying coral-reef textures, and recognizing frog (anuran) calls from audio spectrograms for biodiversity monitoring.

  • pollen
  • coral reefs
  • bioacoustics
  • microscopy
Representative publications
  • J. Strout, B. Rogan, S. M. M. Seyednezhad, M. Bush, E. Ribeiro. “Anuran Call Classification with Deep Learning.” IEEE ICASSP, 2017.
  • A. Daood, E. Ribeiro, M. Bush. “Pollen Grain Recognition Using Deep Learning.” ISVC, 2016.
  • R. Filipovych, A. Daood, E. Ribeiro, M. Bush. “Pollen Recognition in Optical Microscopy by Matching Multifocal Image Sequences.” ICPR, 2016.
  • Z. Ferris, E. Ribeiro, T. Nagata, R. van Woesik. “ReScape: Transforming Coral-Reefscape Images for Quantitative Analysis.” Scientific Reports, 2024.
  • A. Mehta, E. Ribeiro, J. Gilner, R. van Woesik. “Coral Reef Texture Classification Using Support Vector Machines.” VISAPP, 2007.
04

Learning Representations for Complex Signals

Designing and compressing deep models — learning compact, efficient representations through weight pruning and knowledge distillation, and combining information across multiple modalities.

  • deep learning
  • model compression
  • multimodal
Representative publications
  • N. Aghli, E. Ribeiro. “Combining Weight Pruning and Knowledge Distillation for CNN Compression.” IEEE/CVF CVPR Workshops, 2021.
  • N. Aghli, E. Ribeiro. “A Data-Driven Approach to Improve 3D Head-Pose Estimation.” ISVC, 2021.
  • L. Scabini, L. Ribas, E. Ribeiro, O. Bruno. “Deep Topological Embedding with Convolutional Neural Networks for Complex Network Classification.” NetSci-X, 2022.
05

Complex Systems & Network Analysis

Discovering patterns in networked and social data — the temporal rhythms and travelling waves of urban crime, characterization of social-media communities, political influence in multi-party systems, and culture networks.

  • social media
  • crime dynamics
  • culture networks
Representative publications
  • M. Oliveira, E. Ribeiro, C. Bastos-Filho, R. Menezes. “Spatio-temporal Variations in the Urban Rhythm: The Travelling Waves of Crime.” EPJ Data Science, 2018.
  • L. Berrisfeld, E. Ribeiro, R. Menezes. “Estimating Annual Ambient Air Pollution Using Structural Properties of Road Networks.” Environment and Planning B, 2024.
  • J. Faustino, H. Barbosa, E. Ribeiro, R. Menezes. “A Data-Driven Network Approach for Characterization of Political Parties’ Ideology Dynamics.” Applied Network Science, 2019.
  • D. K. Kular, R. Menezes, E. Ribeiro. “Using Network Analysis to Understand the Relation Between Cuisine and Culture.” IEEE Network Science Workshop, 2011.

Funded research

Selected externally funded projects, applying vision and machine learning to problems in aerospace, defense, government, and industry.

  1. National Science Foundation (NSF) 2012–2015$360,000

    Automated Recognition and Mapping of Anuran (Frog) Calls PI

    A full platform for automated species identification and geographic mapping of frog calls — integrating CNN/RNN acoustic classifiers, mobile data collection, and a web-based mapping dashboard for biodiversity monitoring by field biologists.

    Recognition & Classification in Complex Domains
  2. Office of Naval Research (ONR) 2005–2008$250,000

    Automated Image-Based Screening of Marine Coatings PI

    High-throughput image analysis for quantitative assessment of biofouling on novel coating materials, replacing labor-intensive manual scoring — using texture-based segmentation and a calibrated imaging pipeline reproducible across laboratories.

    Recognition & Classification in Complex Domains
  3. Embraer, Inc. 2019–2021$150,000

    AI System for Passenger Identification and Behavioral Analysis PI

    An end-to-end deep-learning system for estimating passenger identity and emotional state from aircraft cabin cameras — using multi-task CNN architectures, domain adaptation for varied lighting, and model compression for edge inference on embedded hardware.

    Learning Representations for Complex Signals
  4. NASA 2024–2025$94,000

    Computer-Vision Pipeline for Rocket-Fuel Slosh Analysis PI

    A GPU-accelerated vision pipeline for automated detection, segmentation, and quantification of fluid-slosh dynamics from high-speed video inside rocket fuel tanks, fusing vision measurements with IMU sensor data against known tank geometry.

    Temporal Understanding of Visual Data

Future directions

Current focus

  • Multimodal perception
  • Efficient deep learning & model compression
  • AI for environmental monitoring

Open challenges

  • Learning from limited data
  • Integrating geometry and deep learning
  • Modeling long-term temporal structure

The throughline: multimodal data contains hidden structure, and vision and machine learning let us discover it — enabling discovery across science and society. See my publications for the full record.