Research

Research

Autonomous drone perception and navigation, robust semantic understanding, and efficient mobile vision.

Research Trajectory

How My Projects Connect

Efficient Vision System Building

VPPlus Efficient video parameter selection for edge analytics
OPA Efficient DNN selection for diverse deployments
Efficient & practical deployment
Autonomous UAV Perception
Showcase Application
SSS Autonomous Drone Delivery to Door
WHAT is in the view?

Semantic Segmentation (SS)

RSSS Robust SS for drone delivery across various environments
FollowMe Annotation Tool
HOW FAR is it?

Metric Geometry (Depth Estimation)

Depth4VTOL Metric depth for vertical UAV motion
AnchorDepth General UAV Motion
Other Application
A4AWD Aerial Weed Detection
2026

AnchorDepth · Fast Metric Depth Estimation with Anchors for Drones

Overview of AnchorDepth for real-time metric depth estimation

Building on Depth4VTOL, AnchorDepth extends real-time metric depth estimation from vertical flight to more general drone movements. Instead of relying on computationally expensive dense multi-view correspondences, it uses stationary, sufficiently large, and geometrically regular objects as anchors to recover metric scale from relative depth. This enables temporally consistent metric depth estimation in real time across common drone flight scenarios.

IEEE/RSJ IROS 2026
2026

Depth4VTOL · Precise Depth Estimation for Vertical Take-off and Landing UAVs

Overview of Depth4VTOL for metric depth estimation during UAV vertical take-off and landing

Beyond semantic understanding, autonomous drone delivery also requires accurate geometric perception of the surrounding environment. Depth4VTOL focuses on metric depth estimation during vertical take-off and landing over residential houses, where predominantly vertical motion provides limited viewpoint diversity. It exploits local geometric changes induced by vertical motion to recover metric scale and achieves real-time depth estimation at 12.4 fps.

ACM MobiCom 2026 · ImmerCom Workshop
2026

FollowMe · Efficient Image Annotation Across Environments

Overview of FollowMe for efficient image annotation

Building on our prior work on robust semantic understanding across environments (RSSS), FollowMe focuses on efficient image annotation. It leverages high-quality annotations from one environment to automatically annotate images of the same physical scene under different times, seasons, and weather conditions, requiring only lightweight human review and correction. FollowMe reduces human effort by 94% while maintaining an average F1 score of 0.9.

ACM MobiCom 2026 · HumanSys Workshop
2025

RSSS · Robust Structural Semantic Segmentation for Autonomous Drone Delivery to Door

Overview of Robust Structural Semantic Segmentation (RSSS)

Building on SSS, RSSS improves the robustness of semantic segmentation for autonomous drone delivery across changing seasons, weather, illumination, and other environmental conditions. Its core idea is to “Let Strong Help Weak”: reliable semantic results obtained under favorable conditions are leveraged to refine weaker predictions in adverse environments using the largely invariant structure and spatial layout of houses.

IEEE/RSJ IROS 2025
2024

SSS · Towards Autonomous Drone Delivery to Your Door Over House-Aware Semantics

Overview of SSS for autonomous drone delivery to the door

We tackle the last-hundred-feet problem for autonomous drone delivery using a semantic segmentation-based approach that progressively guides the drone toward a convenient and safe drop-off point. We leverage single-family house structure to streamline and enhance semantic understanding in the drop-to-door setting.

We were honored to receive the Best Demo Award at ACM HotMobile 2024.

Other Drone Applications

A4AWD · Aerial Weed Detection over Residential Lawns

Overview of A4AWD for UAV-based weed detection over residential lawns

A4AWD targets UAV-based weed detection over residential lawns, where weeds appear extremely small and visually similar to surrounding turfgrass. It introduces a lightweight inference-time augmentation tailored for this challenging green-on-green aerial setting and was evaluated over 165 residential houses.

ICUAS 2026

Earlier Systems

OPA · One-Predict-All for Efficient Deployment

OPA workflow

Instead of training a specialized DNN for each deployment scenario, OPA uses one shallow subnet to test the water. The shallow subnet helps accelerate the search for an appropriate deep subnet, enabling more efficient DNN deployment across diverse use scenarios.

IEEE INFOCOM 2023

VPPlus · Video Processing for Live Video Analytics at the Edge

VPPlus compresses video content live streamed from the device to the edge while preserving the accuracy and timeliness of video analytics. It enlarges the configuration space available for on-device processing and automatically generates feedback to guide joint tuning across more than eight parameters, such as brightness and saturation.

IEEE/ACM IWQoS 2022