# TAN: Temporal Aggregation Network for Dense Multi-label Action Recognition

## TL;DR

The paper introduces TAN, a deep hierarchical network that efficiently captures multi-scale spatio-temporal features for dense multi-label action recognition, outperforming existing methods on key datasets.

## Contribution

TAN decomposes 3D convolutions into spatial and temporal blocks, reducing complexity and improving multi-label action recognition accuracy.

## Key findings

- Outperforms state-of-the-art on Charades by 5%.
- Outperforms state-of-the-art on Multi-THUMOS by 3%.
- Efficient multi-scale spatio-temporal feature aggregation.

## Abstract

We present Temporal Aggregation Network (TAN) which decomposes 3D convolutions into spatial and temporal aggregation blocks. By stacking spatial and temporal convolutions repeatedly, TAN forms a deep hierarchical representation for capturing spatio-temporal information in videos. Since we do not apply 3D convolutions in each layer but only apply temporal aggregation blocks once after each spatial downsampling layer in the network, we significantly reduce the model complexity. The use of dilated convolutions at different resolutions of the network helps in aggregating multi-scale spatio-temporal information efficiently. Experiments show that our model is well suited for dense multi-label action recognition, which is a challenging sub-topic of action recognition that requires predicting multiple action labels in each frame. We outperform state-of-the-art methods by 5% and 3% on the Charades and Multi-THUMOS dataset respectively.