On JRDB, geometry plus optical flow alone reaches 84.3% accuracy for classifying walking/standing/sitting pairs, outperforming the same pipeline augmented with a frozen visual backbone, and transfers zero-shot to a lawnmower platform.
Social scene understand- ing: End-to-end multi-person action localization and col- lective activity recognition
1 Pith paper cite this work, alongside 258 external citations. Polarity classification is still indexing.
1
Pith paper citing it
258
external citations · OpenAlex
fields
cs.RO 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
A Pairwise Human-Human Interaction Detection and Recognition Framework for Mobile Service Robots
On JRDB, geometry plus optical flow alone reaches 84.3% accuracy for classifying walking/standing/sitting pairs, outperforming the same pipeline augmented with a frozen visual backbone, and transfers zero-shot to a lawnmower platform.