Muvienefr Eval

muvienefr-eval

Multi-task View Synthesis with Neural Radiance Fields — Zheng et al. (2023) (arXiv:2309.17450, 2023)

What this evaluates

Evaluates a model's ability to perform multi-task view synthesis by predicting multiple scene properties (RGB, surface normals, shading, edges, keypoints, semantic segmentation) from novel viewpoints, given a set of source-view annotations and camera poses.

Datasets

• Replica — total ?; splits: training (-1), testing (-1)
• SceneNet RGB-D — total ?; splits: training (-1), testing (-1)

Metrics

• RGB (primary) — range: dB
  • Peak Signal-to-Noise Ratio (PSNR) in dB. Higher is better (↑).
• SN — range: other
  • Surface Normal L1 error. Lower is better (↓).
• SH — range: other
  • Shading L1 error. Lower is better (↓).
• ED — range: other
  • Edge Detection L1 error. Lower is better (↓).
• KP — range: other
  • Keypoint L1 error. Lower is better (↓).
• SL — range: [0, 1]
  • Semantic Labeling mIoU. Higher is better (↑).

Input / output format

Input: A set of V source-view task annotations (RGB and K other task annotations) with corresponding camera poses, plus a target camera pose.

Output: Predicted task annotations (RGB, surface normals, shading, edge, keypoint, semantic segmentation) for the target view.

Scoring recipe

Compute per-pixel or per-image metrics between predicted and ground-truth annotations:
- RGB: PSNR = 10 * log10(255^2 / MSE)
- SN, SH, ED, KP: L1 loss = mean(|pred - gt|)
- SL: mIoU = mean(class IoU)
Average across all test scenes and views.

Common pitfalls

• Confusing Inference Setting I (source-view annotations available at test time) with Setting II (RGB-only input at test time).
• Evaluating on seen training scenes instead of strictly unseen novel scenes, which inflates performance.
• Not normalizing or aligning predictions and ground truth for geometric tasks (SN, SH, ED, KP) before computing L1 error.

Evidence (verbatim from paper)

During the evaluation, $\Phi$ is supposed to be generalized to novel scenes that are not seen during training. Table 1: Averaged performance of MuvieNeRF on Replica and SceneNet RGB-D datasets on both training scenes and testing scenes. Metrics reported: RGB (↑), SN (↓), SH (↓), ED (↓), KP (↓), SL (↑).

Citation

@misc{zheng2023muvienefr,
  title={Multi-task View Synthesis with Neural Radiance Fields},
  author={Zheng et al. (2023)},
  year={2023},
  note={arXiv:2309.17450}
}

• arXiv: 2309.17450