Segformer++: Efficient Token-Merging Strategies for High-Resolution Semantic Segmentation

TL;DR

This paper introduces Segformer++, a method that employs token merging strategies to improve the efficiency of transformer-based semantic segmentation on high-resolution images, enabling faster inference and reduced memory use without retraining.

Contribution

It explores various token merging strategies within Segformer, demonstrating significant inference speed improvements while preserving segmentation accuracy on multiple datasets.

Findings

Achieves 61% inference acceleration on Cityscapes without re-training.

Maintains mIoU performance despite token merging.

Facilitates deployment of transformers on resource-constrained devices.

Abstract

Utilizing transformer architectures for semantic segmentation of high-resolution images is hindered by the attention's quadratic computational complexity in the number of tokens. A solution to this challenge involves decreasing the number of tokens through token merging, which has exhibited remarkable enhancements in inference speed, training efficiency, and memory utilization for image classification tasks. In this paper, we explore various token merging strategies within the framework of the Segformer architecture and perform experiments on multiple semantic segmentation and human pose estimation datasets. Notably, without model re-training, we, for example, achieve an inference acceleration of 61% on the Cityscapes dataset while maintaining the mIoU performance. Consequently, this paper facilitates the deployment of transformer-based architectures on resource-constrained devices and in real-time applications.