Sextant
A lightweight spherical angular localization module inspired by the principle of the sextant navigation instrument. It predicts the spherical center azimuth and elevation by reusing frozen visual query vectors from a detector, requiring approximately 0.035 million parameters. The module is plug-and-play and does not require an additional image encoder.
Timeline (last 90 days)
Sextant is an angular localization module inspired by the sextant, which directly measures spherical angles to localize pedestrians.
Sextant has approximately 0.035M (35,000) parameters, is plug-and-play, reuses the frozen detector without changing tracking identities, and requires no additional image encoder.
Sextant takes the visual query vectors output by the frozen detector as input and trains a minimal regression head to predict the spherical center azimuth, elevation angle, and angular range.
In PanoPed-S testing, Sextant improved the HOTA metric of the strongest baseline MOTIP from 47.30 to 49.49, with gains across all eight test sequences.
Without fine-tuning on real data, the localization head trained on synthetic data improved HOTA for MOTIP and HAT on real video by 0.96 to 1.14.
Among comparable systems that do not add a localization image encoder, HAT+Sextant achieved the best score.
All Facts (6)
Sextant is an angular localization module inspired by the sextant, which directly measures spherical angles to localize pedestrians.
50%UnverifiedSextant has approximately 0.035M (35,000) parameters, is plug-and-play, reuses the frozen detector without changing tracking identities, and requires no additional image encoder.
50%UnverifiedSextant takes the visual query vectors output by the frozen detector as input and trains a minimal regression head to predict the spherical center azimuth, elevation angle, and angular range.
50%UnverifiedIn PanoPed-S testing, Sextant improved the HOTA metric of the strongest baseline MOTIP from 47.30 to 49.49, with gains across all eight test sequences.
50%UnverifiedWithout fine-tuning on real data, the localization head trained on synthetic data improved HOTA for MOTIP and HAT on real video by 0.96 to 1.14.
50%UnverifiedAmong comparable systems that do not add a localization image encoder, HAT+Sextant achieved the best score.
50%