qhjqhj00 bat-event-flow-eval
BAT: Learning Event-based Optical Flow with Bidirectional Adaptive Temporal Correlation — Gangwei Xu et al. (2025) (arXiv:2503.03256, 2025)
What this evaluates
Evaluates the accuracy and robustness of event-based optical flow estimation models. It probes the model's ability to predict dense 2D motion fields from sparse, asynchronous event streams, handling varying temporal resolutions and occlusions.
Datasets
- DSEC-Flow — total 8586; splits: train (8170), test (416)
- MVSEC — total ?; splits: train (-1), test (-1)
Metrics
EPE(primary) — range: other- Mean Euclidean distance between predicted and ground truth flow vectors per pixel.
1PE— range: percent- Percentage of ground truth pixels where the end-point error exceeds 1 pixel.
%Out— range: percent- Percentage of ground truth pixels where EPE exceeds 3% or 5% of the ground truth flow magnitude.
Input / output format
Input: Spatiotemporal event streams between two timestamps, typically formatted as event tensors (e.g., B time bins of polarity images) or raw event lists.
Output: Dense 2D optical flow field (u, v vectors) for every pixel in the target frame.
Scoring recipe
def evaluate(pred_flow, gt_flow):
epe = np.mean(np.linalg.norm(pred_flow - gt_flow, axis=-1))
npe_1 = 100 * np.mean(np.linalg.norm(pred_flow - gt_flow, axis=-1) > 1.0)
mag_gt = np.linalg.norm(gt_flow, axis=-1)
out_3 = 100 * np.mean(epe > 0.03 * mag_gt)
out_5 = 100 * np.mean(epe > 0.05 * mag_gt)
return {'EPE': epe, '1PE': npe_1, '%Out_3%': out_3, '%Out_5%': out_5}
Common pitfalls
- MVSEC evaluation splits are fixed to outdoor_day2 (train) and outdoor_day1 (test); using other sequences breaks comparability.
- NPE metrics (1PE, 2PE, 3PE) measure the percentage of pixels where the Euclidean flow error exceeds a pixel threshold, not the angular error or component-wise error.
Evidence (verbatim from paper)
The end-point error (EPE) is the primary metric used to evaluate the accuracy of optical flow predictions for both DSEC-Flow and MVSEC. Additionally, we report the percentage of ground truth pixels with an optical flow magnitude error greater than N (NPE), where N is either 1, 2, or 3, as well as the Angular Error (AE) for DSEC-Flow. For MVSEC, we report the percentage of ground truth pixels with EPE exceeding 3 and 5% of the flow magnitude (%Out).
Citation
@misc{xu2025bat,
title={BAT: Learning Event-based Optical Flow with Bidirectional Adaptive Temporal Correlation},
author={Gangwei Xu et al. (2025)},
year={2025},
note={arXiv:2503.03256}
}
- arXiv: 2503.03256