Zijun Long

Vision atlas

Selected projects in computer vision and multimodal learning.

RoboEye first ranks candidate objects using 2D features, then refines ambiguous matches using 3D keypoints.
RoboEyeRobotic object identification

ICRA 2026

Identifying what a robot picks up

RoboEye first ranks objects by their appearance in 2D. When similar candidates are hard to tell apart, it checks whether their 3D keypoints match.

RAD matches test patches against a memory of normal features at several layers to produce an anomaly heatmap.
RADRetrieval-based anomaly detection

ICML 2026

Finding defects by looking up normal examples

RAD compares local features of a test image with those of defect-free examples. Regions without a close match are flagged for inspection.

ADaFuSE combines a dialogue and a generated image using an adaptive gate and semantic experts to retrieve a target image.
ADaFuSEInteractive text-to-image retrieval

SIGIR 2026

Searching with words and generated images

A generated image can capture a searcher’s intent or introduce distracting details. ADaFuSE learns how much weight to give each image and text cue.

Selected publications

All publications →