Publications by Year

For publications before 2017, please visit my Google Scholar page.

Featured Publications

Here are the featured publications. For the complete list, please visit my Google Scholar page.

CLIP

What CLIP Knows but Cannot Say: Recovering Negation from Frozen Intermediate Features

2026 New

ECCV 2026 Long Oral — one of only 28 long orals chosen from 10,473 submissions (top 0.27% of all submissions, top 1% of accepted papers)

AdaLLaVA

Learning to Inference Adaptively for Multimodal Large Language Models

2025

IEEE/CVF International Conference on Computer Vision (ICCV), 2025

Details

Proposes an inference-time policy that picks per-query settings—image resolution, token budget, modality routing, early-exit—to trade accuracy, latency, and energy on the fly. Yields higher throughput at fixed utility and graceful quality under tight budgets across representative MLLMs.

Acceptance Rate: 24.0% (2,699 / 11,239)

Keywords: adaptive inference, MLLM, dynamic modality routing, early exit, token budgeting, latency-energy trade-offs.

skald

SKALD: Learning-Based Shot Assembly for Coherent Multi-Shot Video Creation

2025

IEEE/CVF International Conference on Computer Vision (ICCV), 2025

Details

Introduces a learned clip-coherence objective that scores spatial/temporal/semantic flow and a beam-search assembler to select coherent shot sequences under style/length constraints. Outperforms rule-based and caption-similarity baselines while remaining plug-and-play with common editors/generators.

Acceptance Rate: 24.0% (2,699 / 11,239)

Keywords: video assembly, coherence, shot selection, beam search, generative editing.

SWSeg

Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity

2025

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Details

Aligns labeled/unlabeled features via Sliced-Wasserstein distance and adds a uniformity prior to avoid cluster collapse, boosting mIoU under extreme label sparsity and class imbalance. Compatible with modern backbones and stable across confidence thresholds.

Acceptance Rate: 22.1% (2,878 / 13,008)

Keywords: semi-supervised segmentation, feature alignment, Sliced-Wasserstein, uniformity, class imbalance.

agile3d

Agile3D: Adaptive Contention- and Content-Aware 3D Object Detection for Embedded GPUs

2025

ACM International Conference on Mobile Systems, Applications, and Services (MobiSys)

Right-sizes LiDAR detection by reacting jointly to scene content and on-device contention, adjusting resolution/blocks to hold tail latency and energy in check with minimal AP loss. Demonstrates portable gains across embedded platforms with a lightweight runtime.

Acceptance Rate: 18.0% (42 / 233)

Keywords: 3D detection, embedded GPU, adaptive scheduling, contention-aware, LiDAR.

recon

ReCON: Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories

2024

European Conference on Computer Vision (ECCV)

Details

Speeds diffusion generation by reusing retrieved concept trajectories—preserving fidelity and style without retraining—and complements existing schedulers. Works across diverse prompts and models with consistent step reductions.

Acceptance Rate: 27.9% (2,395 / 8,585)

Keywords: diffusion, retrieval-augmented generation, prompt trajectories, speed-ups.

SENSORBFT: Energy-Efficient Fault-Tolerant Target Localization

2024

IEEE International Conference on Distributed Computing Systems (ICDCS)

Details

Uses lightweight consensus with sensing to localize under Byzantine faults, lowering message cost and energy while maintaining accuracy. Scales to dense deployments with bounded overhead.

Acceptance Rate: 21.9% (121 accepted; ≈553 submissions)

Keywords: federated sensing, Byzantine resilience, consensus, energy efficiency.

Adaptive Virtuoso

Adaptive Object Detection algorithms for Resource Constrained Autonomous Robotic Systems

2024

SPIE Defense + Commercial Sensing, Proc. SPIE Vol. 13058, Disruptive Technologies in Information Sciences VIII (SPIE DCS) — Collaboration with the Army Research Lab stemming from internship work; includes a working software artifact

Details

Extends Virtuoso for real-time detection under contention by parallelizing detection and tracking and using multi-resolution tracker initialization matched to box size—maintains throughput even with costlier detectors. Validated on embedded hardware.

Keywords: robotics, multi-threaded vision, tracker initialization, real-time.

Asgard

Asgard: Are NoSQL databases suitable for ephemeral data in serverless workloads?

2023

Frontiers in High Performance Computing (Frontiers HPC) — Joint work with an undergraduate and an alum now at Salesforce; perspective-style article with early results on “serverless cloud computing meets NoSQL” for volatile workloads

Details

Empirical study of NoSQL stores for short-lived, bursty serverless data paths; analyzes throughput-latency-cost under TTL/ephemeral patterns. Provides configuration guidance for scaling transient state without overprovisioning.

Keywords: serverless, ephemeral data, NoSQL, cost-performance, TTL.

DSN DISrupt

How to Learn Collaboratively - Federated Learning to Peer-to-Peer Learning and What's At Stake

2023

IEEE/IFIP International Conference on Dependable Systems and Networks — DISrupt Workshop (DSN-DISrupt) — “Disruptive paper” outlining our hub-and-spoke decentralized learning approach: higher availability and privacy than federated learning (FL) and lower compute/communication overhead than peer-to-peer (P2P) learning—best of both worlds

Details

Maps FL ↔ P2P trade-offs (privacy, personalization, robustness, communication cost) and proposes deployment metrics and design patterns. Highlights hybrid topologies (hub-and-spoke + peer gossip) and reliability/auditing gaps.

Keywords: federated learning, peer-to-peer, collaboration, privacy, robustness.

Vega

Vega: Drone-based Multi-Altitude Target Detection for Autonomous Surveillance

2023

IEEE International Conference on Distributed Computing in Sensor Systems and Internet of Things (DCOSS-IoT)

Details

Fuses complementary views from different flight altitudes and schedules sensing to balance coverage, energy, and detection confidence. Improves robustness to occlusions/scale changes under onboard compute limits.

Acceptance Rate: 23% (20 / 86)

Keywords: UAV sensing, multi-altitude fusion, energy-aware scheduling, surveillance.

FLAIR

FLAIR: Defense against Model Poisoning Attack in Federated Learning

2023

ACM Asia Conference on Computer and Communications Security (AsiaCCS)

Details

Combines anomaly scoring with robust aggregation to down-weight poisoned updates while preserving benign personalization. Sustains accuracy across heterogeneous clients and attack strengths with modest overhead.

Acceptance Rate: 18% (418 / 2,322)

Keywords: federated learning, poisoning defense, robust aggregation, anomaly detection.

virtuoso

Virtuoso: Energy- and Latency-Aware Streamlining of Streaming Videos on SOCs

2022

ACM Transactions on Design Automation of Electronic Systems (ACM TODAES)

Details

Enforces energy SLOs by jointly scheduling operators, resolutions, and model variants to keep latency within budget on mobile SoCs; includes deployable runtime and open implementation. Evaluations show stable QoE under tight power caps.

Keywords: mobile SoC, energy SLO, operator scheduling, video pipelines.

smartadapt

SMARTADAPT: Multi-branch Object Detection Framework for Videos on Mobiles

2022

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Details

Switches among detection branches based on content dynamics, improving accuracy-latency trade-offs for on-device video analytics. Shows gains across varied motion/texture regimes.

Acceptance Rate: 25.3% (2,064 / 8,161)

Keywords: adaptive detection, mobile vision, content-aware routing, on-device.