Language: English | 繁體中文

Research Vision 研究主軸

Advanced Computer Vision Lab — Assured Computer Vision: Lean, Autonomous, Broad-Spectrum

As generative AI blurs the boundary between authentic and fabricated media, autonomous systems demand vision that never fails silently, and Earth observation enters a data-rich new era, the bar for deployable visual intelligence keeps rising. ACVLab responds with four interlocking research pillars.

Assured Visual Intelligence ensures that every visual AI output can be trusted — whether detecting DeepFakes under heavy compression, defending against adversarial perturbations, or authenticating media through proactive watermarking — providing the accountability that forensic, medical, and regulatory settings require.

Efficient Computing Systems rethink computation at every level of abstraction: prefix-scan reformulations of exact attention (ELSA), bitstream-level forensics that skip pixel decoding entirely, adaptive quantization that preserves accuracy at ultra-low bit widths (QuantTune/FracQuant), and joint transmission-restoration for bandwidth-constrained satellites — cutting latency, memory, and energy cost for sustainable, real-time deployment.

Autonomous Visual Perception extends vision from 2D images into 3D physical space: material-aware scene reconstruction with hyperspectral unmixing, BEV adversarial defense for self-driving (BFDM), physics-aligned shadow and reflection removal that feeds robust features to downstream robotic pipelines (PhaSR, ReflexSplit), and uncertainty-aware 3D annotation for autonomous driving datasets.

Broad-Spectrum Scientific Sensing pushes perception beyond the visible: universal hyperspectral restoration via vision-language prompts (PromptHSI), real-time CubeSat compressed sensing recognized with the Future Technology Award, hyperspectral pansharpening through sparse spectral representations (S³RNet), and cross-spectral forgery detection that reveals manipulation invisible to RGB analysis.

These pillars do not operate in isolation. Hyperspectral forensics merges trust with spectral sensing. On-satellite real-time inference merges efficiency with broad-spectrum data. BEV adversarial defense merges trust with embodied perception. This cross-pillar synergy is not accidental — it reflects a single underlying conviction: deployment-grade visual intelligence must be simultaneously trustworthy, efficient, embodied, and perceptually complete.

Adjacent computational direction. Beyond the four visual pillars, we also explore lean computational systems for emerging workloads such as resource-aware quantum simulation. This is an extension of the lab's deployment-oriented systems perspective, not a fifth visual pillar.

Research viewAll cards remain visible; the selected option changes priority/order only. Within each priority group, cards remain newest first.

Order / priority

Main ResearchCompetition & WorkshopEfficient Computing Systems

Research Pillars

Autonomous Visual Perception: PhaSR, ReflexSplit, autonomous driving, tracking, embodied perception, 3D reconstruction
Assured Visual Intelligence: GRACEv2, UMCL, DDD-Net, DeepFake detection, proactive authentication, trustworthy media analysis
Broad-Spectrum Scientific Sensing: PromptHSI, S³RNet, CubeSat compressed sensing, remote sensing, satellite imaging
Efficient Computing Systems: ELSA, QuantTune, FracQuant, bitstream-level inference, CubeSat on-board processing, edge deployment

A short introduction to my research: [PDF] (Latest updated: Oct. 2024)

Main Research

Visual Restoration & Reconstruction

Robust Shadow Removal

PhaSR: Generalized Image Shadow Removal with Physically Aligned Priors

Accepted to IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026.

Shadow removal under complex and multi-source lighting is hindered by the mismatch between physical illumination priors and learned features. PhaSR couples physically aligned normalization with geometry-semantic rectification to deliver robust shadow removal that generalizes beyond traditional single-light settings.

Research Direction. Autonomous Visual Perception / Robust Scene Recovery

[arXiv] [GitHub]

Main Research

Visual Restoration & Reconstruction

Reflection Separation in the Wild

ReflexSplit: Single Image Reflection Separation via Layer Fusion-Separation

Accepted to IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026.

Reflections on glass introduce nonlinear layer mixing that often breaks existing separation networks. ReflexSplit uses dual-stream fusion-separation blocks and curriculum training to achieve robust performance on both synthetic and real-world benchmarks.

Research Direction. Autonomous Visual Perception / Robust Scene Recovery

[arXiv] [GitHub]

Efficient Computing Systems

Efficient AI

Efficient AI Inference

ELSA: Exact Linear-Scan Attention for Fast and Memory-Light Vision Transformers

Accepted to CVPR 2026 Findings (CVPRF).

ELSA reformulates exact softmax attention as a prefix scan over an associative monoid, achieving memory-light inference with provable FP32 stability and no retraining. Implemented in Triton and CUDA C++, it improves deployability on both data-center and edge hardware.

Research Direction. Efficient Computing Systems / Hardware-Agnostic Inference

[arXiv] [GitHub]

Efficient Computing Systems

Quantum Simulation

Memory-Bounded Quantum Simulation

LC-Implicit-QAOA: Active-Workspace-Capped Exact Objective-and-Gradient Evaluation for Training over Bounded QUBO Light Cones

Preprint on arXiv.

LC-Implicit-QAOA evaluates exact QUBO-QAOA objectives and shared-parameter gradients directly from bounded causal cones, materializing neither a global state vector nor a global cost table, so cost follows the largest neighbourhood instead of the qubit count. It profiles the problem and commits to a declared workspace budget before allocation; on the reported 3-regular benchmarks it reaches n=28 in 45 ms using 12 MB, while a matched state-plus-cost simulator runs out of memory, with reported comparisons of 28× faster than NVIDIA CUDA-Q at n=28 and 31× faster than CUAOA at n=26.

Research Direction. Efficient Computing Systems / Resource-Aware Quantum Simulation

[arXiv] [GitHub]

Efficient Computing Systems

Quantum Simulation

Per-Instance Selection for Quantum Simulation

RASP-QAOA: Resource-Aware Per-Instance Selection for Exact QAOA Simulation

Preprint on arXiv.

RASP-QAOA treats simulator selection as a complete executable action—representation, adapter, precision mode and memory policy fixed together—and removes every action that cannot implement the requested semantics or fit the memory budget before ranking what remains. In a content-disjoint 60-request H200 evaluation it succeeds on all 31 requests with an admissible action, compared with 19 for a fixed best-known backend, and reports a 25× lower failure-penalized PAR10 score; a depth-1 stump matches gradient boosting, indicating that the gain comes from describing the options rather than from model capacity.

Research Direction. Efficient Computing Systems / Resource-Aware Quantum Simulation

[arXiv] [GitHub]

Efficient Computing Systems

Efficient AI

Candidate-Aware Early Exit for Diffusion Language Models

Where and When to Commit: Candidate-Aware Decoding for Diffusion Language Models

Preprint on arXiv.

LATCH separates when to stop an entire diffusion-language-model generation from where to accelerate block commitments. Confidence-Verified Commit (CVC) checks confidence and sustained stability over the dynamically extracted candidate span, while Block-Wise Early Commit (BWEC) accelerates only non-final blocks. Across 11 tasks and 22 evaluation settings on LLaDA and Dream, accuracy stays within 2.0 percentage points of full decoding, with 9.3–17.8× end-to-end TPS speedups on short-answer tasks and 2.0–3.3× on long-reasoning tasks.

Research Direction. Efficient Computing Systems / Efficient AI

[arXiv] [GitHub]

Competition & Workshop

Visual Restoration & Reconstruction

Illumination-Invariant Semantic Priors

CANDLE: Illumination-Invariant Semantic Priors for Color Ambient Lighting Normalization

Published in IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) 2026.

CANDLE uses semantic priors to normalize color ambient lighting while preserving scene structure. The paper reports a 1.22 dB PSNR improvement and 3rd place in the NTIRE 2026 ALN Color Lighting track plus 2nd place in the White Lighting fidelity track.

Research Direction. Autonomous Visual Perception / Robust Illumination Normalization

[CVF Open Access] [arXiv] [GitHub]

Competition & Workshop

Medical & Biomedical Vision

Rare-Pathology Video Capsule Endoscopy

VISTA: Validation-Guided Integration of Spatial and Temporal Foundation Models with Anatomical Decoding for Rare-Pathology VCE Event Detection

Post-competition preprint on arXiv.

VISTA combines EndoFM-LV and DINOv3 with validation-guided fusion and anatomy-aware temporal decoding for event-level rare-pathology detection in video capsule endoscopy. The authors report hidden-test mAP@0.5 of 0.3530 and mAP@0.95 of 0.3235; post-competition global threshold refinement reaches 0.3726 and 0.3431, respectively, and is reported as second place for Team ACVLab.

Research Direction. Broad-Spectrum Scientific Sensing / Medical Video Intelligence

[arXiv] [RARE-VISION challenge]

Efficient Computing Systems

Efficient AI

Quantization-Friendly Deployment

QuantTune: Optimizing Model Quantization with Adaptive Outlier-Driven Fine Tuning

Published in IEEE International Conference on Multimedia Information Processing and Retrieval (MIPR) 2025.

QuantTune addresses outlier-driven dynamic range amplification during Transformer quantization and substantially reduces accuracy loss under low-bit settings. The method requires no extra inference-time hardware complexity and transfers across ViT, BERT, and OPT models.

Research Direction. Efficient Computing Systems / Quantization-Aware Deployment

[arXiv] [IEEE Xplore]

Main Research

Hyperspectral & Scientific Sensing

Universal Hyperspectral Restoration

PromptHSI: Universal Hyperspectral Image Restoration with Vision-Language Modulated Frequency Adaptation

Published in IEEE Transactions on Geoscience and Remote Sensing (TGRS), Early Access, Feb. 2026.

PromptHSI is a universal all-in-one framework for hyperspectral restoration that combines frequency-aware modulation with vision-language guided prompt learning. A single model can handle cloud occlusion, blur, noise, and spectral band loss across remote sensing scenarios.

Research Direction. Broad-Spectrum Scientific Sensing / Hyperspectral Restoration

[IEEE Xplore] [arXiv] [GitHub]

Main Research

Trustworthy Media & DeepFake

Media Security & DeepFake Robustness

Towards Robust DeepFake Detection under Unstable Face Sequences: Adaptive Sparse Graph Embedding with Order-Free Representation and Explicit Laplacian Spectral Prior

Submitted to IEEE Transactions on Information Forensics and Security (TIFS).

GRACEv2 targets unstable face sequences caused by compression, occlusion, and shuffled or missing frames. By combining order-free temporal graph embedding with an explicit Laplacian spectral prior, it improves robust DeepFake detection under severe real-world disruptions.

Research Direction. Assured Visual Intelligence / Robust DeepFake Detection

[arXiv]

Main Research

Trustworthy Media & DeepFake

Cross-Compression DeepFake Detection

UMCL: Unimodal-Generated Multimodal Contrastive Learning for Cross-compression-rate Deepfake Detection

Published in International Journal of Computer Vision (IJCV), Jan. 2026.

UMCL synthesizes compression-robust multimodal cues, including rPPG, temporal landmarks, and semantic embeddings, from a single visual input. The framework improves cross-compression DeepFake detection while preserving interpretable feature relationships.

Research Direction. Assured Visual Intelligence / Cross-Compression Forensics

[Springer] [DOI] [arXiv]

Competition & Workshop

Medical & Biomedical Vision

COVID-19 Symptoms Detection in CT Scan

Selected challenge papers and results

IEEE ECCV Workshop 2022 [1st place in COV19D challenge]

Spatial-Slice Feature Learning using Visual Transformer and Essential Slices Selection Module for COVID-19 Detection of CT Scans in the Wild

IEEE ICCV Workshop 2021 [3rd place in COV19D challenge]

Adaptive Distribution Learning with Statistical Hypothesis Testing for COVID-19 CT Scan Classification

Our models are designed for noisy, in-the-wild CT scans and remain robust across varying spatial and slice resolutions.

Competition & Workshop

Autonomous Perception & Tracking

Semantic Segmentation for Autonomous Driving (2021-)

Selected papers for robust and efficient scene understanding

IEEE ICME Workshop 2022

Augmented-Training-Aware Bisenet for Real-Time Semantic Segmentation [PDF]

IEEE ICASSP 2022

DCSN: Deformable Convolutional Semantic Segmentation Neural Network for Non-Rigid Scenes [PDF]

These projects focus on stable, real-time semantic understanding for autonomous driving, balancing robustness and low-compute deployment.

Main Research

Hyperspectral & Scientific Sensing

Deep Compressed Sensing for Hyperspectral Images (2020-)

Selected papers for efficient satellite sensing

IEEE Transactions on Geoscience and Remote Sensing

DCSN: Deep Compressed Sensing Network for Efficient Hyperspectral Data Transmission of Miniaturized Satellite [PDF]

CVGIP 2020

Deep Joint Compression and Super-Resolution Low-Rank Network for Fast Hyperspectral Data Transmission

[Project] [GitHub]

以深度學習為基礎之高光譜 / 多光譜影像超解析度與壓縮感知技術開發。

Main Research

Autonomous Perception & Tracking

Decision-Making of Autonomous Vehicles Using Vision Information (2019-)

Selected work on robust visual decision-making

Multimedia Tools and Applications

Deep Learning-based Vehicle Trajectory Prediction based on Generative Adversarial Network for Autonomous Driving Applications

IEEE ICCE-TW 2020

Learning to Predict Risky Driving Behaviors for Autonomous Driving

[Large-Scale Vehicle Collision Dataset @ TW] [Link]

自駕車視覺系統之危險駕駛行為預測與台灣道路地區資料庫建置。

Chih-Chung Hsu (許志仲)

Research Vision 研究主軸

Research Pillars

Robust Shadow Removal

Reflection Separation in the Wild

Efficient AI Inference

Memory-Bounded Quantum Simulation

Per-Instance Selection for Quantum Simulation

Candidate-Aware Early Exit for Diffusion Language Models

Illumination-Invariant Semantic Priors

Rare-Pathology Video Capsule Endoscopy

Quantization-Friendly Deployment

Universal Hyperspectral Restoration

Media Security & DeepFake Robustness

Cross-Compression DeepFake Detection

New SOTA SR Model

Semi-Supervised Learning in CT Scan Detection

Ultra Fast Hyperspectral Image Compressive Sensing

COVID-19 Symptoms Detection in CT Scan

Social Media Prediction as Longitudinal Task (2022-)

Semantic Segmentation for Autonomous Driving (2021-)

Fake Image/Video (DeepFake) Detection (2018-)

Deep Compressed Sensing for Hyperspectral Images (2020-)

Decision-Making of Autonomous Vehicles Using Vision Information (2019-)

Social Media Prediction (2016-)

Identity-Preserving Face Hallucination (2018-2020)

Large-Scale Image Clustering (2016-2017)

Image Deblocking and Super-Resolution (2013-2014)

Super-Resolution of Textured Video (2012-2014)

Quality Assessment for Image Retargeting (2011-2013)

Super-Resolution (2010-2011)

Face Hallucination (2008-2010)

Video Forensics (2007-2008)

Image Authentication (2006-2007)

You missed

[Award] We just received the 2026 Future Tech Award (未來科技獎)!!

[News!] I joined ITRI’s Smart Robotics Innovation and Application R&D Center as Principal Researcher!

[Award] We won 2nd Place in the RARE-VISION post-competition evaluation!

[News!] Five more papers accepted to CVPR 2026 Workshops!

Counter