This paper explores how humans rapidly recognize daily activities from complex visual inputs. We utilized videos of 16 daily activities from the MoVi dataset to compare three movement analysis strategies:
- Temporal Movement Primitives (TMPs): Decomposes movements into weighted sums of temporally smooth basis functions.
- Legendre Polynomial Coefficients: Projects joint-coordinate trajectories onto an orthogonal polynomial basis.
- Autoencoder Latent Embeddings.
Results indicated that Legendre coefficients and TMPs achieved the highest classifier accuracy, followed by autoencoders. We identified two key features for movement classification:
- General Posture: The average spatial configuration that distinguishes different activities.
- Nine Critical Joints: These joints were the most predictive for movement classification.
Interestingly, high classification accuracy did not guarantee effective movement reconstruction. TMPs preserved temporal dynamics, resulting in perceptually natural motion, while reconstructions from Legendre coefficients retained only the average posture, appearing frozen.
These findings reveal how movement information is organized: the static body configuration suffices to classify activities, but temporal dynamics are essential for reconstructing their unfolding. This distinction clarifies the features the visual system may rely on for rapid action recognition, suggesting that postural features could enable efficient movement screening in clinical applications, while dynamic information remains crucial for movement generation.
Blogger's Review: This paper provides a compelling comparative analysis of movement classification strategies, shedding light on the differing roles of posture and dynamics, which has significant implications for action recognition and generation in various applications.