[Proposal] Motion-adaptive temporal frame sampler for training dataset

[AyushChauhan910]

Summary

I've implemented a motion-adaptive temporal frame sampler for the
RandomUniformSampler in training/dataset/vos_sampler.py. It replaces
uniform stride with motion-density-proportional frame budget allocation,
selecting more frames from high-motion intervals during training.

Problem

The current RandomUniformSampler treats all temporal positions as equally
informative. Analysis of 15 DAVIS-2017 sequences shows this causes
systematic under-sampling of high-motion transitions — exactly the frames
where object appearance changes most rapidly and boundary learning is
most critical.

Solution

AdaptiveTemporalSampler (sam2/utils/adaptive_sampler.py):

• Scores frames via lightweight L1 pixel-diff (subsamples every 4th frame)
• Allocates budget_ratio of frame budget to high-motion regions
• Falls back to uniform sampling on any exception
• Fully backward-compatible: opt-in via sampler_cfg in config

Measured Results (15 DAVIS-2017 val sequences, 8 frames/clip)

Metric	Uniform	Adaptive	Delta
Mean high-motion coverage	0.122	0.127	+0.005
High-motion coverage (bear)	0.12	0.25	+0.13
High-motion coverage (boat)	0.10	0.19	+0.09
Frames per clip	8	8	0

5/15 sequences show clear improvement. Sequences with evenly-distributed
motion (bus, car-shadow) correctly receive near-uniform selection — the
sampler adapts to clip content rather than forcing dense sampling everywhere.

Full retraining would be needed to measure downstream J&F improvement,
which is outside scope of this PR.

Implementation Status

• AdaptiveTemporalSampler implemented (pure PyTorch + PIL, no OpenCV)
• Integrated into vos_sampler.py (backward-compatible)
• Hydra config support (sampler_cfg.type: adaptive)
• 8 unit tests passing (CPU-only)
• Motion analysis tooling added

Happy to open a PR if this direction is welcome. I can also run ablation
over different motion_threshold values if that would help the review.