ConvMLP: Hierarchical Convolutional MLPs for Vision

TL;DR

ConvMLP introduces a hierarchical, lightweight convolutional MLP architecture that enhances visual recognition tasks, including object detection and segmentation, with competitive accuracy and efficiency.

Contribution

It proposes a novel stage-wise convolutional MLP design that improves applicability to downstream tasks and reduces computational cost compared to prior MLP-based models.

Findings

Achieves 76.8% top-1 accuracy on ImageNet-1k with fewer parameters and MACs.

Demonstrates competitive transfer learning performance on detection and segmentation.

Offers a lightweight, stage-wise design suitable for various vision tasks.

Abstract

MLP-based architectures, which consist of a sequence of consecutive multi-layer perceptron blocks, have recently been found to reach comparable results to convolutional and transformer-based methods. However, most adopt spatial MLPs which take fixed dimension inputs, therefore making it difficult to apply them to downstream tasks, such as object detection and semantic segmentation. Moreover, single-stage designs further limit performance in other computer vision tasks and fully connected layers bear heavy computation. To tackle these problems, we propose ConvMLP: a hierarchical Convolutional MLP for visual recognition, which is a light-weight, stage-wise, co-design of convolution layers, and MLPs. In particular, ConvMLP-S achieves 76.8% top-1 accuracy on ImageNet-1k with 9M parameters and 2.4G MACs (15% and 19% of MLP-Mixer-B/16, respectively). Experiments on object detection and semantic segmentation further show that visual representation learned by ConvMLP can be seamlessly transferred and achieve competitive results with fewer parameters. Our code and pre-trained models are publicly available at https://github.com/SHI-Labs/Convolutional-MLPs.