SNN-DEVO checkpoints

Trained patch-selector checkpoints from the master's thesis Exploring Neuromorphic Front-Ends for Event-Based SLAM (Ekansh Bajpai, Paderborn University, 2026).

Code: ekansh-bajpai/snn-devo. Each checkpoint belongs to one branch of that repository and only loads correctly with that branch checked out, because the patch selector is chosen by SCORER_TYPE in constants.py.

All models were trained on TartanEvents with the same schedule; the checkpoints are taken at step 120000, the step used for every evaluation in the thesis.

File	Patch selector	Code branch	Code tag (evaluation)	SCORER_TYPE
v1-cnn-baseline/v1-cnn-baseline_step120000.pth	DEVO CNN, retrained on TartanEvents	v1-cnn-baseline	v1-cnn-baseline-eval	CNN
v2-snn-voxel/v2-snn-voxel_step120000.pth	SNN, temporal-voxel encoding	v2-snn-voxel	v2-snn-voxel-eval	TEMPORAL
v2.1-snn-voxel-scaled/v2.1-snn-voxel-scaled_step120000.pth	SNN, temporal-voxel encoding + BatchNorm, full BPTT, learnable affine score head	v2.1-snn-voxel-scaled	v2.1-snn-voxel-scaled-eval	TEMPORAL
v3-snn-timesurface/v3-snn-timesurface_step120000.pth	SNN, time-surface encoding	v3-snn-timesurface / main	v3-snn-timesurface-eval	TIME_SURFACE

The original DEVO-pretrained CNN (DEVO.pth) is not hosted here; get it from the DEVO authors with ./download_model.sh in the code repository.

Usage

Access is gated: request access on this page and log in with hf auth login before downloading.

Set up the code repository as described in its README, then:

git checkout v3-snn-timesurface        # branch matching the checkpoint
hf download upb-rat-ekansh/snn-devo v3-snn-timesurface/v3-snn-timesurface_step120000.pth --local-dir checkpoints

python -m evals.eval_evs.eval_mvsec_evs_ts --datapath=<MVSECPATH> \
    --weights=checkpoints/v3-snn-timesurface/v3-snn-timesurface_step120000.pth \
    --val_split=splits/thesis/mvsec_eval.txt --trials=5

Run the evaluation as a module from the repository root. v1, v2 and v2.1 use the voxel modules (evals.eval_evs.eval_{tartan,mvsec,fpv}_evs); v3 uses the _ts variants. Checksums are in SHA256SUMS (sha256sum -c SHA256SUMS).

Citation

@mastersthesis{bajpai2026neuromorphic,
  title  = {Exploring Neuromorphic Front-Ends for Event-Based SLAM},
  author = {Bajpai, Ekansh},
  school = {Paderborn University},
  year   = {2026}
}

@inproceedings{klenk2023devo,
  title     = {Deep Event Visual Odometry},
  author    = {Klenk, Simon and Motzet, Marvin and Koestler, Lukas and Cremers, Daniel},
  booktitle = {International Conference on 3D Vision, 3DV 2024},
  pages     = {739--749},
  publisher = {IEEE},
  year      = {2024}
}

License

MIT, same as the code. DEVO is MIT-licensed by the TUM Computer Vision Group and Princeton Vision & Learning Lab.