Aaron Chen.Let's talk

RESEARCH / 2021 — 2026

Research with
a clearer vision.

Restoring the details. Understanding the whole.
Explore the ideas, methods, and results behind my work.

Jump to all 13 publications
SELECTED RESEARCH STORIESSCROLL TO EXPLORE ↓

Previews play silently while visible. Use the video controls to pause or resume.

01 / 05CVPR 2026
Image Generation and EditingMultimodal Learning

Restore. Assess. Repeat.

Restoration that knows when to look again.

Real images rarely arrive with just one kind of degradation. RAR brings image restoration and quality assessment into a shared, iterative process.

Inside the research

The approach

Identify degradation, restore the image, and verify its quality within a unified latent-space framework.

Why it matters

A more adaptive way to handle unknown and composite image degradations.

Full paper title

Restore, Assess, Repeat: A Unified Framework for Iterative Image Restoration

Paper Project Watch video
Iterative quality assessment and restoration: recovering detail from a degraded image.

Iterative quality assessment and restoration: recovering detail from a degraded image.

Research preview · Source · View full figure
02 / 05CVPR 2026
Multimodal LearningVisual Understanding

RobustVisRAG

Reliable answers start with resilient perception.

Blur, noise, and shadows can disrupt how a model retrieves and reads visual documents. RobustVisRAG separates meaningful content from visual degradation.

Inside the research

The approach

Use a causality-guided dual-path framework to disentangle semantic information and distortion signals.

Why it matters

More dependable retrieval and grounded generation from imperfect visual evidence.

Full paper title

RobustVisRAG: Causality-Aware Vision-Based Retrieval-Augmented Generation under Visual Degradations

Paper Project Watch video
Retrieve
Retrieving the relevant visual document despite degraded image quality.
Generation
Comparing generated answers to a question about a degraded retrieved document.
Research preview · Source · View full figure
03 / 05CVPR 2025 / Highlight
Image Generation and EditingMultimodal LearningVisual Understanding

UniRestore

Better to our eyes. Better for the task.

A visually pleasing image is not always the most useful one for a vision model. UniRestore connects perceptual quality with downstream task needs.

Inside the research

The approach

Adapt a diffusion prior with complementary feature restoration and task-aware feature fusion.

Why it matters

One restoration framework designed for both human perception and machine understanding.

Full paper title

UniRestore: Unified Perceptual and Task-Oriented Image Restoration Model Using Diffusion Prior

Paper Project Watch video
Image restoration for human perception and downstream vision tasks.

Image restoration for human perception and downstream vision tasks.

Research preview · Source · View full figure
04 / 05ICCV 2025
Domain GeneralizationVisual Understanding

PDAF

Understanding scenes beyond familiar conditions.

A segmentation model should keep making sense of the world when the environment changes. PDAF models the hidden shifts between visual domains.

Inside the research

The approach

Use a latent domain prior and probabilistic diffusion alignment to refine feature representations.

Why it matters

A path toward more generalizable semantic segmentation in unfamiliar urban scenes.

Full paper title

Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation

Paper Project Watch video
A visual comparison of baseline and PDAF semantic segmentation predictions.

A visual comparison of baseline and PDAF semantic segmentation predictions.

Research preview · Source · View full figure
05 / 05ECCV 2024
Visual Understanding

APGCC

Every point deserves a better match.

Counting a crowd is also a matching problem. APGCC gives point-based models clearer guidance on where to look and which proposals to trust.

Inside the research

The approach

Combine Auxiliary Point Guidance with Implicit Feature Interpolation to stabilize proposal-target matching.

Why it matters

More stable learning for crowd counting and localization across varied scenes.

Full paper title

Improving Point-based Crowd Counting and Localization Based on Auxiliary Point Guidance

Paper Project Watch video
Crowd counting and localization with point detections in a real scene.

Crowd counting and localization with point detections in a real scene.

Research preview · Source · View full figure

THE RESEARCH INDEX

All publications.

13 PAPERS / 2021–2026
Sequential image restoration examples from Restore, Assess, Repeat
CVPR 2026

Restore, Assess, Repeat: A Unified Framework for Iterative Image Restoration

I-Hsiang Chen and collaborators · First author

Image Generation and EditingMultimodal Learning
Paper Project Watch video
RobustVisRAG overview showing visual document retrieval and grounded answer generation
CVPR 2026

RobustVisRAG: Causality-Aware Vision-Based Retrieval-Augmented Generation under Visual Degradations

I-Hsiang Chen and collaborators · First author

Multimodal LearningVisual Understanding
Paper Project Watch video
UniRestore comparison of perceptual and task-oriented restoration of a street scene
CVPR 2025✳ Highlight

UniRestore: Unified Perceptual and Task-Oriented Image Restoration Model Using Diffusion Prior

I-Hsiang Chen and collaborators · First author

Image Generation and EditingMultimodal LearningVisual Understanding
Paper Project Watch video
PDAF framework for latent domain modeling and domain compensation in semantic segmentation
ICCV 2025

Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation

I-Hsiang Chen and collaborators · First author

Domain GeneralizationVisual Understanding
Paper Project Watch video
APGCC architecture showing auxiliary point guidance for crowd counting and localization
ECCV 2024

Improving Point-based Crowd Counting and Localization Based on Auxiliary Point Guidance

I-Hsiang Chen and collaborators · First author

Visual Understanding
Paper Project Watch video
Research figure for semantic guidance in non-homogeneous image dehazing
CVPRW 2023

Semantic Guidance Learning for High-Resolution Non-Homogeneous Dehazing

I-Hsiang Chen and collaborators · Co-author

Image Generation and EditingVisual Understanding
Paper
TSRFormer two-stage refinement for single-image shadow removal
CVPRW 2023

TSRFormer: Transformer Based Two-Stage Refinement for Single Image Shadow Removal

I-Hsiang Chen and collaborators · Co-author

Image Generation and Editing
Paper
RVSL research figure for vehicle similarity learning in hazy scenes
ECCV 2022

RVSL: Robust Vehicle Similarity Learning in Real Hazy Scenes Based on Semi-supervised Learning

I-Hsiang Chen and collaborators · First author

Visual UnderstandingDomain Generalization
Paper Code
SJDL-Vehicle joint defogging and vehicle re-identification framework
AAAI 2022

SJDL-Vehicle: Semi-supervised Joint Defogging Learning for Foggy Vehicle Re-identification

I-Hsiang Chen and collaborators · First author

Visual UnderstandingDomain Generalization
Paper Code
Missing Recovery single-image reflection removal research figure
IEEE TIP 2022

Missing Recovery: Single Image Reflection Removal Based on Auxiliary Prior Learning

I-Hsiang Chen and collaborators · Co-author

Image Generation and Editing
Paper
Single-image desnowing examples from ALL Snow Removed
ICCV 2021

ALL Snow Removed: Single Image Desnowing Algorithm Using Hierarchical Dual-tree Complex Wavelet Representation and Contradict Channel Loss

I-Hsiang Chen and collaborators · Co-author

Image Generation and Editing
Paper Code
DesmokeNet image smoke removal research figure
IEEE TCSVT 2021

DesmokeNet: A Two-stage Smoke Removal Pipeline Based on Self-Attentive Feature Consensus and Multi-Level Contrastive Regularization

I-Hsiang Chen and collaborators · Co-author

Image Generation and Editing
Paper Code
ContourletNet research figure for single-image rain removal
BMVC 2021

ContourletNet: A Generalized Rain Removal Architecture Using Multi-Direction Representation and Hierarchical Decomposition

I-Hsiang Chen and collaborators · Co-author

Image Generation and Editing
Paper Code

Bibliography preview: author lists are abbreviated. Refer to the linked papers for complete citations; final metadata review is pending.

RESEARCH PRESENTATION

Project video

Loading YouTube player...

Open on YouTube