Efficient Multi-Scale Deformable Attention on GPUs
Where the time actually goes in the slowest operator of DETR-style detectors, and kernels that run it up to fourteen times faster on a fraction of the memory.
2026
2025
Balancing Shared and Task-Specific Representations: A Hybrid Approach to Depth-Aware Video Panoptic Segmentation
One model that segments a street scene, measures how far away everything is, and tracks objects between frames — each better than models built for a single task.