The research question
A detector can respond differently when the same face is captured at a different distance or image scale. I study why a multi-level screen-illumination baseline loses detection evidence as the captured face becomes smaller.
What I am working on
- Reproducing the measurement and decision pipeline, from face regions and optical flow to window-level classification and video-level voting.
- Building matched near/far captures and digital-scaling controls, with original, processing-only, and authorized face-swap conditions.
- Examining the effects of image scale, flow measurements, exposure, and filtering rules separately.
A controlled initial batch
The initial batch contains four matched near/far recording pairs at 30 cm and 60 cm. Three processing conditions per recording give 24 evaluated video conditions. This is a small pilot, not a study of 24 independent participants.
| Condition | Near · 30 cm | Far · 60 cm |
|---|---|---|
| Original | 4/4 Real | 4/4 Real |
| Processing-only | 4/4 Real | 4/4 Real |
| Authorized face swap | 4/4 Fake | 0/4 Fake |
Counts refer to complete-video decisions on the same initial batch. They describe a failure pattern under these conditions, rather than an overall accuracy estimate.
Current direction
The next stage examines measurement compensation, false-positive control, and filtering rules together. Parameters must be fixed before testing a newly acquired batch, so that same-batch observations can be separated from independent validation.
The work remains in progress. The research summary focuses on the experimental design and the observed baseline failure, with independent generalization still to be established.