Event-Based De-Snowing for Autonomous Drivings

Manasi Muglikar, Nico Messikommer, Marco Cannici, Davide Scaramuzza

Robotics and Perception Group, University of Zürich

IEEE Transactions on Robotics (T-RO) 2026

Code release for the T-RO 2026 paper on event-based image de-snowing for autonomous driving.

Citation

If you use this codebase, or the datasets accompanying the paper, please cite the following publications:

@inproceedings{muglikar26tro,
  title={Event-Based De-Snowing for Autonomous Driving},
  author={Muglikar, Manasi and Messikommer, Nico and Cannici, Marco and Scaramuzza, Davide},
  booktitle={Transactions of Robotics},
  year={2026}
}

Environment

Verified baseline:

• Python 3.6.13
• PyTorch 1.8.0
• torchvision 0.9.0
• CUDA 11.x

Install dependencies:

python -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

Dataset Layout

Training and testing scripts expect each sequence folder to contain:

<sequence>/
  masked_images/
  gt_images/
  events/
  voxel/           # generated by preprocess.py
  voxel_20ms/      # used by test.py for DSEC/real driving eval

For training, point --train_dir and --test_dir to directories containing per-sequence subfolders.

Reproducible Pipeline

1. Preprocess Events

Generate cached voxel grids from event H5 files:

bash preprocess.sh

Expected behavior:

• One background preprocessing job per sequence (up to MAX_JOBS in the script).
• New .npy files created in each sequence voxel folder.

2. Train

Use the provided launcher:

bash train.sh

Or run directly:

python train.py \
  --arch EvSnowNet_Paper \
  --batch_size 2 \
  --gpu 0 \
  --train_ps 256 \
  --lr_initial 0.0001 \
  --save_dir ./logs \
  --warmup \
  --resume \
  --pretrain_weights ./logs/EvSnowNet_Paper/models/model_epoch_40.pth \
  --train_dir /path/to/dataset/train/ \
  --test_dir /path/to/dataset/test/ \
  --save_images \
  --nepoch 91 \
  --checkpoint 10

Expected console output includes lines such as:

• saving to : ./logs/EvSnowNet_Paper
• ===> Loading datasets
• ===> Start Epoch , End Epoch
• Checkpoint saved at epoch !

Artifacts are written to:

• logs//.txt
• logs//models/model_epoch_.pth
• logs//results/train/

3. Inference

Run batch testing over subfolders:

bash test.sh

Or run directly for a single sequence:

python test.py \
  --input_dir /path/to/sequence \
  --result_dir /path/to/output \
  --model_name EvSnowNet_Paper \
  --weights ./logs/EvSnowNet_Paper/models/model_epoch_40.pth \
  --batch_size 1 \
  --has_groundtruth

Expected output structure:

<result_dir>/
  image/
  events/
  gt/
  evsnownet_paper/

4. Metrics

Compute dataset-level PSNR/SSIM:

python evaluate.py

Expected output:

• Per-sequence processing logs.
• Final summary: Dataset Mean PSNR and Dataset Mean SSIM.
• Metrics text file written near the inference root.

Acknowledgement

The code is derived from multiple sources, in particular Snowformer and RLP. Thanks for their great works!