billpsomas/vits_dino_simpool_ep300

Self-supervised ViT-S/16 (small-sized Vision Transformer with patch size 16) model with SimPool

ViT-S model with SimPool (gamma=1.25) trained on ImageNet-1k for 300 epochs. Self-supervision with DINO.

SimPool is a simple attention-based pooling method at the end of network, introduced on this ICCV 2023 paper and released in this repository. Disclaimer: This model card is written by the author of SimPool, i.e. Bill Psomas.

Motivation

Convolutional networks and vision transformers have different forms of pairwise interactions, pooling across layers and pooling at the end of the network. Does the latter really need to be different? As a by-product of pooling, vision transformers provide spatial attention for free, but this is most often of low quality unless self-supervised, which is not well studied. Is supervision really the problem?

Method

SimPool is a simple attention-based pooling mechanism as a replacement of the default one for both convolutional and transformer encoders. For transformers, we completely discard the [CLS] token. Interestingly, we find that, whether supervised or self-supervised, SimPool improves performance on pre-training and downstream tasks and provides attention maps delineating object boundaries in all cases. One could thus call SimPool universal.

Evaluation with k-NN

k	top1	top5
10	72.56	87.638
20	72.434	89.24
100	70.526	90.582
200	69.33	90.424

BibTeX entry and citation info

@misc{psomas2023simpool,
      title={Keep It SimPool: Who Said Supervised Transformers Suffer from Attention Deficit?}, 
      author={Bill Psomas and Ioannis Kakogeorgiou and Konstantinos Karantzalos and Yannis Avrithis},
      year={2023},
      eprint={2309.06891},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}