Related papers: DropPos: Pre-Training Vision Transformers by Reconstructing Dropped Positions

DropPos: Pre-Training Vision Transformers by Reconstructing Dropped Positions

URL: http://arxiv.org/abs/2309.03576v2
Date: Fri, 22 Sep 2023 00:54:47 GMT
Title: DropPos: Pre-Training Vision Transformers by Reconstructing Dropped Positions
Authors: Haochen Wang, Junsong Fan, Yuxi Wang, Kaiyou Song, Tong Wang, Zhaoxiang Zhang
Abstract summary: We present DropPos, a novel pretext task designed to reconstruct Dropped Positions. The code is publicly available at https://github.com/Haochen-Wang409/DropPos.
Score: 63.61970125369834
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Abstract: As it is empirically observed that Vision Transformers (ViTs) are quite insensitive to the order of input tokens, the need for an appropriate self-supervised pretext task that enhances the location awareness of ViTs is becoming evident. To address this, we present DropPos, a novel pretext task designed to reconstruct Dropped Positions. The formulation of DropPos is simple: we first drop a large random subset of positional embeddings and then the model classifies the actual position for each non-overlapping patch among all possible positions solely based on their visual appearance. To avoid trivial solutions, we increase the difficulty of this task by keeping only a subset of patches visible. Additionally, considering there may be different patches with similar visual appearances, we propose position smoothing and attentive reconstruction strategies to relax this classification problem, since it is not necessary to reconstruct their exact positions in these cases. Empirical evaluations of DropPos show strong capabilities. DropPos outperforms supervised pre-training and achieves competitive results compared with state-of-the-art self-supervised alternatives on a wide range of downstream benchmarks. This suggests that explicitly encouraging spatial reasoning abilities, as DropPos does, indeed contributes to the improved location awareness of ViTs. The code is publicly available at https://github.com/Haochen-Wang409/DropPos.