We propose a novel Motion Forecasting method that learns traffic-aware scene graphs to capture the relationships between objects and their motion patterns, achieving state-of-the-art performance on benchmarks.

We propose a novel Image Editing method that leverages intrinsic cues from flow inversion to determine where to edit and what to preserve, achieving state-of-the-art performance on benchmarks.

We propose a novel Environmental Change Detection (ECD) framework that handles misaligned and uncurated reference images by aggregating multiple reference candidates and rich semantic cues, achieving robust and state-of-the-art performance on standard benchmarks.

We propose a novel P2E VPR method that leverages the properties of hyperbolic space to address the matching problem between perspective views and equirectangular images.

We propose BridgeTA, a cost-effective knowledge distillation framework that bridges the representation gap between LiDAR-Camera fusion and camera-only models through a lightweight teacher-assistant network, achieving superior efficiency and performance in BEV map segmentation.

We propose A2LC, a novel active and automated label correction framework for semantic segmentation that integrates automated correction guided by annotator feedback with adaptive sample acquisition, achieving significantly higher efficiency and performance than prior ALC methods.

This paper proposes a real-time RGB-D semantic segmentation method that effectively encodes depth information and fuses it with RGB features to enhance segmentation performance while maintaining computational efficiency.

We propose a novel BEV segmentation network that adaptively utilizes features of various scales based on the location in the BEV space.

We propose a storage-efficient large-scale 3D mapping method that employs a discrete representation based on a decomposition strategy.

This paper proposes a real-time RGB-based street floor segmentation method that identifies traversable and non-traversable curbs to support mobile robot navigation in urban environments.

We propose 3D-DEEP, a deep-learning architecture designed for road scene understanding using elevation patterns derived from disparity-filtered and LiDAR-projected images.