A Spitting Image: Modular Superpixel Tokenization in Vision Transformers

TL;DR

This paper introduces a modular superpixel tokenization method for Vision Transformers that improves interpretability and pixel-level prediction without sacrificing classification accuracy.

Contribution

It presents a content-aware, scale- and shape-invariant tokenization approach that decouples tokenization from feature extraction in ViTs, enhancing interpretability and dense prediction.

Findings

Improves faithfulness of attributions

Enables pixel-level granularity in dense prediction

Maintains classification performance

Abstract

Vision Transformer (ViT) architectures traditionally employ a grid-based approach to tokenization independent of the semantic content of an image. We propose a modular superpixel tokenization strategy which decouples tokenization and feature extraction; a shift from contemporary approaches where these are treated as an undifferentiated whole. Using on-line content-aware tokenization and scale- and shape-invariant positional embeddings, we perform experiments and ablations that contrast our approach with patch-based tokenization and randomized partitions as baselines. We show that our method significantly improves the faithfulness of attributions, gives pixel-level granularity on zero-shot unsupervised dense prediction tasks, while maintaining predictive performance in classification tasks. Our approach provides a modular tokenization framework commensurable with standard architectures, extending the space of ViTs to a larger class of semantically-rich models.