Agglomerative Token Clustering
This addresses the challenge of efficiently reducing token counts in vision models, particularly at low keep rates, offering a practical improvement for researchers and practitioners in computer vision.
The paper tackles the problem of token merging in vision tasks by introducing Agglomerative Token Clustering (ATC), which achieves state-of-the-art performance across image classification, synthesis, and object detection & segmentation without extra parameters, even matching prior methods without fine-tuning.
We present Agglomerative Token Clustering (ATC), a novel token merging method that consistently outperforms previous token merging and pruning methods across image classification, image synthesis, and object detection & segmentation tasks. ATC merges clusters through bottom-up hierarchical clustering, without the introduction of extra learnable parameters. We find that ATC achieves state-of-the-art performance across all tasks, and can even perform on par with prior state-of-the-art when applied off-the-shelf, i.e. without fine-tuning. ATC is particularly effective when applied with low keep rates, where only a small fraction of tokens are kept and retaining task performance is especially difficult.