Issue with running metric on more density captured clouds

[BradenCladock]

📍 Which component is affected?

Objective metric (please specify in description)

🔥 What happened?

The metrics I've been running comfortably on standard PCQA benchmark sequences OOM on the captured data I actually want to evaluate. I can downsample as a workaround, but I'm worried that downsampling itself changes what the metric "sees" relative to what a human subject would see. What's the recommended way to evaluate PCQA on point clouds that exceed the metric's training-set scale?

🪜 Steps to reproduce

*

🖥️ Environment

📊 Dataset and input

No response

📋 Logs / traceback

✅ Pre-flight checklist

• I have searched the existing issues and confirmed this is not a duplicate.
• I have read the relevant section of the README / sub-README.
• I am willing to test a proposed fix and report back.

[Terriao]

Hi, this is a real tension and there isn't a single best answer — the right path depends on what you're trying to claim with the score.
The core problem: PCQA metrics are trained on benchmarks whose point clouds fit in a particular point-count range, and the learned feature scales are implicitly tuned to that range. Inputs outside the training scale produce out-of-distribution feature responses regardless of whether you can fit them in memory. So you have to make it fit in a way that doesn't shift the input distribution away from what the metric was trained on. You can try the following methods to solve this issue.

1. Downsample to training scale.
   Resample both the reference and the distorted clouds to a target point count in the training range. This is fast, OOMs go away, and the metric is now running on inputs it was trained for. The trade-off you're worried about is real, but bounded: the metric is evaluating the perceptual quality of the resampled cloud, which is what the metric is qualified to evaluate.

2. Tile-and-aggregate.
   Partition the cloud spatially (e.g. octree leaves at a chosen depth, or contiguous voxel blocks), run the metric on each tile within VRAM, and aggregate the per-tile scores into a single number. This preserves the full resolution but introduces an aggregation choice that affects the final number, and tiles with few points may produce unreliable per-tile scores that dominate the mean. Works well when the metric is local; works poorly when it relies on global statistics.