Adapt-STFormer

Flexible and Efficient Spatio-Temporal Transformer for Sequential Visual Place Recognition

ICRA 2026

Idan Lau, Chao Chen, Ge Jin, Chen Feng,
New York University
arXiv

Overview

Computationally efficient and sequence length-adaptable visual place recognition by combining spatial and temporal understanding in one recurrent module.


Overview Image

How It Works

Adapt-STFormer's Recurrent-DTE module considers previous image features as latent states along with the current image, accounting for the temporal relationship across image sequences with variable length. This unifies the modeling of spatio-temporal relationships across images in one module

Method

Overview Image

1. Image Encoding: For each image in the input sequence, we use CCT384 to obtain a set of image tokens. This effectively precomputes the inputs to the Recurrent-DTE module.
2. Recurrence: Given a new image, the DTE treats its CCT tokens as both key and value inputs, while the DTE outputs from the previous image serves as the query for the current iteration. For initialization, a learnable positional embedding is added to the first image's tokens to become the DTE query.
3. Aggregation: We concatenate the DTE outputs for each image along the batch dimension and apply the learnable mean pooling SeqGeM to flatten across the batch dimension. This result is fed to SeqVLAD for aggregation into the final descriptor.

BibTeX



Coming Soon

Acknowledgements

Chen Feng is the corresponding author. This work was supported in part through the NYU IT High Performance Computing resources, services, and staff expertise. We also thank Somik Dhar for early contributions.