I2e Event Classification Eval

i2e-event-classification-eval

I2E: Real-Time Image-to-Event Conversion for High-Performance Spiking Neural Networks — Ruichen Ma et al. (arXiv:2511.08065, 2025)

What this evaluates

Evaluates the classification performance of Spiking Neural Networks trained on synthetic event streams generated from static images, and tests the transferability of these models to real-world neuromorphic sensor data.

Datasets

• I2E-CIFAR10 — total ?; splits: train (-1), test (-1)
• I2E-CIFAR100 — total ?; splits: train (-1), test (-1)
• I2E-ImageNet — total ?; splits: train (-1), test (-1)
• CIFAR10-DVS — total ?; splits: train (-1), test (-1)

Metrics

• Accuracy (primary) — range: percent
  • Standard classification accuracy: the proportion of correctly predicted class labels out of the total number of test instances, reported as a percentage.

Input / output format

Input: Event streams represented as either dense boolean tensors or sparse coordinate lists, generated from resized static images (224x224 for ImageNet, 128x128 for CIFAR) or captured by real DVS sensors.

Output: Discrete class labels corresponding to the image categories.

Scoring recipe

correct = 0
for pred, gold in zip(predictions, gold_labels):
    if pred == gold:
        correct += 1
accuracy = (correct / len(gold_labels)) * 100

Common pitfalls

• Confusing synthetic I2E-generated event data with real DVS sensor recordings, which have different noise and sparsity characteristics.
• Overlooking the impact of data augmentation strategies (Baseline-I vs Baseline-II) on performance, as applying standard augmentations to source images before conversion drastically boosts accuracy.
• Assuming fixed timestep counts are optimal; the paper shows performance varies significantly with timestep order and count, requiring careful ablation.

Evidence (verbatim from paper)

On I2E-ImageNet, MS-ResNet34 (Baseline-II) reaches 60.50% accuracy, surpassing the best prior result on other event-based ImageNet datasets by over 8%. The dramatic performance increase from Baseline-I to Baseline-II across all datasets demonstrates that I2E is not only capable of generating high-quality event data but also uniquely enables the modern training strategies required to unlock the full potential of deep SNNs.

Citation

@misc{ma2025i2e,
  title={I2E: Real-Time Image-to-Event Conversion for High-Performance Spiking Neural Networks},
  author={Ruichen Ma et al.},
  year={2025},
  note={arXiv:2511.08065}
}

• arXiv: 2511.08065