Vision Transformers have received significant attention due to their impressive performance in many vision tasks. While the token mixer or attention block has been studied in great detail, the channel mixer or feature mixing block (FFN or MLP) has not been explored in depth albeit it accounts for a bulk of the parameters and computation in a model. In this work, we study whether sparse feature mixing can replace the dense connections and confirm this with a block diagonal MLP structure that improves the accuracy by supporting larger expansion ratios. To improve the feature clusters formed by this structure and thereby further improve the accuracy, a lightweight, parameter-free, channel covariance attention (CCA) mechanism is introduced as a parallel branch during training. This design of CCA enables gradual feature mixing across channel groups during training whose contribution decays to zero as the training progresses to convergence. This allows the CCA block to be discarded during inference, thus enabling enhanced performance with no additional computational cost. The resulting $\textit{Scalable CHannEl MixEr}$ (SCHEME) can be plugged into any ViT architecture to obtain a gamut of models with different trade-offs between complexity and performance by controlling the block diagonal structure size in the MLP. This is shown by the introduction of a new family of SCHEMEformer models. Experiments on image classification, object detection, and semantic segmentation, with different ViT backbones, consistently demonstrate substantial accuracy gains over existing designs, especially under lower FLOPs regimes. For example, the SCHEMEformer establishes a new SOTA of 79.7% accuracy for ViTs using pure attention mixers on ImageNet-1K at 1.77G FLOPs.
Action localization networks are often structured as a feature encoder sub-network and a localization sub-network, where the feature encoder learns to transform an input video to features that are useful for the localization sub-network to generate reliable action proposals. While some of the encoded features may be more useful for generating action proposals, prior action localization approaches do not include any attention mechanism that enables the localization sub-network to attend more to the more important features. In this paper, we propose a novel attention mechanism, the Class Semantics-based Attention (CSA), that learns from the temporal distribution of semantics of action classes present in an input video to find the importance scores of the encoded features, which are used to provide attention to the more useful encoded features. We demonstrate on two popular action detection datasets that incorporating our novel attention mechanism provides considerable performance gains on competitive action detection models (e.g., around 6.2% improvement over BMN action detection baseline to obtain 47.5% mAP on the THUMOS-14 dataset), and a new state-of-the-art of 36.25% mAP on the ActivityNet v1.3 dataset. Further, the CSA localization model family which includes BMN-CSA, was part of the second-placed submission at the 2021 ActivityNet action localization challenge. Our attention mechanism outperforms prior self-attention modules such as the squeeze-and-excitation in action detection task. We also observe that our attention mechanism is complementary to such self-attention modules in that performance improvements are seen when both are used together.