MegaEvent - Multi-viewpoint Geo-localization with Event Cameras

This repository contains the evaluation and training code for MegaEvent, our state-of-the-art visual place recognition system for event cameras.

ArXiv Paper • Pre-trained models • Springfield-Event-VPR dataset

Install

Install Pixi, clone this repository, then run:

pixi install

Linux and Windows resolve a CUDA 12 build of PyTorch; Apple Silicon uses MPS when available, otherwise CPU. Run the test suite with pixi run test.

Retrieve

pixi run megaevent retrieve --reference reference.h5 --query query.h5 --top-k 5 --output results

EventCV reads the event files. A file is sliced into 50 ms windows (--window-ms); a directory supplies one sample per event file. --output receives retrievals.csv with query_id,reference_id,rank,score, alongside samples.json and run.json describing the windows and the settings used. Add --save-previews 10 to also write match images for ten sampled queries into previews/.

Evaluate

pixi run megaevent retrieve --reference reference.h5 --query query.h5 --ground-truth positives.npy --output results

Ground truth is a boolean [reference, query] matrix aligned to the windows, or a CSV with query_id,reference_id columns; pass --gt-layout query-reference for the transposed matrix. Evaluation adds Recall@1/5/10/20 and writes metrics.json.

Models

Retrieval downloads the weights it needs on the first run, so there is no separate download step. --model megaevent_vits14 (default) or --model megaevent_vitb14 fetches a pinned revision from AdamHines/megaevent and caches it in src/ckpts; later runs reuse the cached file.

Training code lives in src/megaevent/training with recipes in configs/train/, and is not documented yet.

Citation

If you use our work, please cite the following paper:

@misc{hines2026megaevent,
      title={Multi-viewpoint Geo-localization with Event Cameras}, 
      author={Adam D. Hines and Michael Milford and Tobias Fischer},
      year={2026},
      eprint={2609.21219},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.21219}, 
}