Our research lies at the intersection of data-efficient learning, human–AI interaction, and 3D vision.
We build intelligent vision systems that adapt to new environments, interact seamlessly with humans, and generalize across diverse modalities.
Training segmentation AI requires dense pixel-level annotations, which are extremely expensive and labor-intensive.
Weakly supervised segmentation overcomes this limitation by learning from cheaper and more widely available weak labels (such as image-level tags, points, boxes, or scribbles).
We build novel frameworks that leverage weak labels, reducing the costs while maintaining high-quality segmentation.
We further investigate ways to harness the general knowledge of foundation models to strengthen weak supervision.
AI suffers performance drop when deployed in new environments different from the training data (known as domain shift).
TTA tackles this by allowing the model to adjust itself on-the-fly during inference, without access to any other supervision.
We devise various TTA methods regarding 3D understanding, especially for autonomous driving and indoor scene perception.
Understanding the 3D world is essential for intelligent systems that operate in real environments (as we are living in 3D!)
3D vision aims to perceive and interpret geometric structures from diverse inputs such as point clouds or multi-view images.
We study full spectrum of 3D vision (perception & generation) with applications to autonomous driving and physical AI.
We are particularly interested in combining 3D vision with multi-modal data (such as 2D vision, language, and audio).
Automated segmentation struggles in practice, as it is not always clear which object to segment or how detailed it should be.
Interactive segmentation addresses this issue by using simple user inputs to guide the system.
With this interaction, the model predicts a precise segmentation mask following the user's intention.
We explore methods to adapt deployed AI models to make them better suited to each situation and context.
Training modern AI systems often requires a large amount of labeled data, which is expensive and time-consuming to collect.
Active learning aims to identify which data is the most informative, and then asks the human "oracle" to label only those.
By repeating this process, the model improves quickly while requiring far fewer labeled examples.
We investigate active learning framework for multi-modal tasks (2D–3D–Language) to maximize learning efficiency.
We also explore scenarios where the "oracle" is not a human but another AI, pioneering new directions for scalable learning.
