SAM-H uses SAM 2 masks for training-free homography tracking, +18.4pp on PlanarTrack

ducha_aiki · x · 2026-09-11

An ECCV 2026 paper by Jonas Serych and Jiri Matas (CTU Prague) repurposes modern segmentation trackers for planar object tracking.

• SAM-H: a training-free pipeline that estimates 8-DoF homography poses directly from segmentation mask contours. Applied to SAM 2 masks, it sets a new state of the art on the PlanarTrack benchmark by a wide margin (+18.4pp on p@5).
• WOFTSAM: shows segmentation-based and correspondence-based homography estimation are complementary, outperforming all prior methods on both PlanarTrack and POT-210.
• The authors also provide precise re-annotations of PlanarTrack initial poses for more accurate high-precision benchmarking.

Code and re-annotations are open-sourced.

Original post →

More from Research

• NVIDIA Open-Sources BioNeMo Inference Runtime to GPU-Accelerate Protein Models — AllThingsApx · 2026-09-12
• 31 million protein complex predictions run on NVIDIA BioNeMo, saving an estimated 1.35 GWh — AllThingsApx · 2026-09-12
• ApprenticeBench: Agents Continually Learn Real Jobs, Surpassing Human Pros — ysu_nlp · 2026-09-12
• Why RL Environments Work Better in 2026: Greenblatt's Two Reasons — dejavucoder · 2026-09-12
• Signal65 launches PINNACLE, an agentic AI benchmark scoring correct work over raw throughput — ryanshrout · 2026-09-12
• EvoHarnessBench: adding tools to agents can silently degrade abilities they already had — mohitban47 · 2026-09-11

Research channel →