Research · Public overview

Video modelsEvidence

A collaborative video-analysis system followed by controlled evaluation of representations, external evidence, and operating behaviour under domain shift.

Role
Collaborative research & evaluation direction
Setting
Video analysis under domain shift
Focus
Data integrity and external evaluation
Publication
TransFlower, accepted for oral presentation at KES 2026
01 · The question

Does a promising model change improve the evidence that matters?

Video systems can rank benchmark clips well while still confusing ordinary interactions with the target event. I worked across a collaborative video model and a later evaluation campaign to separate model behaviour, data exposure, domain shift, and operating decisions.

The work moved from reproducing a functioning baseline to checking what was actually being evaluated, then testing frozen representations with external and benign-behaviour checks.

02 · The approach

Verify the comparison, measure the trade-off, keep the rejection condition visible.

  1. 01 · Verify

    Make model and data identity inspectable.

    I established model identity, evaluation conditions, and the comparison boundary before interpreting a model difference.

  2. 02 · Evaluate

    Measure ranking and benign behaviour together.

    I directed controlled comparisons that treated external evidence, operating trade-offs, and content integrity as part of the research question.

  3. 03 · Revise

    Let failed checks change the next experiment.

    Promising development signals were carried forward as hypotheses only when their limitations and possible failure modes remained explicit.

03 · Qualitative findings

Useful evidence has more than one axis.

  1. 01 · Baseline

    A clean replay is a prerequisite for a useful comparison.

    Model identity, checkpoint fidelity, content exposure, and selection boundaries determine what a measured change can mean.

  2. 02 · External evidence

    A positive signal can remain exploratory.

    Frozen-feature readouts showed a research signal under a separate camera cohort, while limited sites and grouped uncertainty kept the conclusion bounded.

  3. 03 · Operating behaviour

    Better ranking does not settle the false-alarm question.

    Candidate readouts were judged alongside benign clips, because a useful system needs acceptable behaviour at the operating point as well as discrimination.

04 · Contribution

From model adaptation to experiment design.

Within the collaborative TransFlower project, I was responsible for much of the Transformer/VideoMAE implementation and carried out the documented experiments. My work also included developing the RGB/optical-flow I3D component and integrating the combined inference path. In the later campaign, I set research priorities, directed controlled comparisons, reviewed conflicting analyses, inspected flagged media, and decided which hypotheses to retain.

The later work used pretrained V-JEPA representations for task-specific readouts and evaluation.

05 · Reviewer access

The detailed evidence is available for invited review.

Reviewer access includes the detailed report, aggregate figures, and the limits that shape their interpretation.

06 · Starting points

Collaborative work, upstream models, careful evaluation.