Adversarial Multi-view Networks for Activity Recognition
summary
Proposes DAMUN: a multi-view CNN extractor that builds temporal, spatial and spatio-temporal representations, fuses them with a trainable Hadamard fusion module, and uses a Siamese adversarial discriminator to remove subject-specific variance. Evaluated with iterative left-one-subject-out on MHEALTH, PAMAP2 and UCIDSADS, the model achieves the best average performance versus baselines, finds the SST+MST+TCS view trio most useful, shows consistent gains from adversarial training, and surfaces a persistent “class collapse” on hard-to-recognize activities.
AI-written summary · about
chapters (7)
- 0:00Start
- 0:53Formulation
- 1:25Problems
- 2:43Siamese Adversarial
- 3:31Datasets
- 4:19Overall Comparison
- 4:34Findings
related talks
06mUnsupervised Human Activity Representation Learning with Multi-task Deep Clustering
06mDANA: Dimension-Adaptive Neural Architecture for Multivariate Sensor Data
28mLarge Multimodal Models: Understanding the State of the Art in Compute
28mDeep Learning-Based Human Activity Recognition for Physical Skills...
26mAccelerating Explorations in Vision and Multimodal AI Using Pytorch...
01h 30mDeep Learning Techniques for Monocular Depth Estimation
06mMulti-Head Spatio-Temporal Attention Mechanism For Urban Anomaly Event Prediction
34mComputer Vision and NLP for Multi-Task Fashion Attrb. Modeling