Kitti Optical Flow Eval

kitti-optical-flow-eval

Occlusion Aware Unsupervised Learning of Optical Flow From Video — Jianfeng Li et al. (arXiv:2003.01960, 2020)

What this evaluates

Evaluates the accuracy of unsupervised optical flow estimation methods on standard driving scenes. It measures the pixel-wise displacement error between predicted and ground-truth flow fields to quantify estimation quality.

Datasets

• KITTI2012 — total ?; splits: train (40864), val (3822)

Metrics

• EPE (primary) — range: other
  • Average Euclidean distance between estimated and ground-truth optical flow vectors across all pixels.

Input / output format

Input: Consecutive video frames (e.g., frame t and frame t+1) resized to 832x256.

Output: A 2-channel optical flow field representing horizontal and vertical displacement per pixel.

Scoring recipe

def compute_epe(pred_flow, gt_flow):
    diff = pred_flow - gt_flow
    epe = np.mean(np.sqrt(diff[..., 0]**2 + diff[..., 1]**2))
    return epe

Common pitfalls

• The paper does not specify whether EPE is computed over the entire image or only valid ground-truth pixels, which can significantly affect reported values.
• The method is trained unsupervised using photometric loss but evaluated on supervised KITTI benchmarks, creating a potential train-test distribution mismatch.

Evidence (verbatim from paper)

For evaluation, we use end-point error (epe) which is defined as the average Euclidean distance between estimated and ground-truth optical flows.

Citation

@misc{li2020occlusion,
  title={Occlusion Aware Unsupervised Learning of Optical Flow From Video},
  author={Jianfeng Li et al.},
  year={2020},
  note={arXiv:2003.01960}
}

• arXiv: 2003.01960