| --- |
| language: en |
| license: mit |
| tags: |
| - gaze-estimation |
| - knowledge-distillation |
| - pytorch |
| - mediapipe |
| - reproducibility |
| metrics: |
| - angular-error |
| --- |
| |
| # LIPE gaze estimation: models and reproducibility archive |
|
|
| This Hugging Face repository contains LIPE source code, model checkpoints, processed V16 |
| participant files, and the evidence required to reproduce the knowledge-distillation |
| implementation audit. |
|
|
| Raw MPIIGaze and Gaze360 images are not distributed. Users must obtain those datasets |
| under their original licenses when rebuilding preprocessing or clean teacher caches. |
|
|
| The archived KD comparison is reproducible, but it is not an isolated causal KD ablation: |
| the historical KD-on runs also changed the hard loss, and the recovered teacher pipeline |
| contained loading, angle-decoding, and coordinate-alignment defects. The corrected study |
| therefore uses matched control, point-KD, oracle-gated KD, and shuffled-teacher arms. |
|
|
| Start with [REPRODUCING_KD_INVESTIGATION.md](REPRODUCING_KD_INVESTIGATION.md). The current |
| evidence and safe paper wording are in: |
|
|
| - `artifacts/kd-teacher-trap-diagnostic/REPORT.md` |
| - `artifacts/kd-teacher-trap-diagnostic/CLEAN_KD_EXPERIMENT_SPEC.md` |
| - `artifacts/kd-teacher-trap-diagnostic/PAPER_KD_REWRITE_DRAFT.md` |
|
|
| MPIIGaze leave-one-participant-out evaluation is described as subject-independent LOPO, |
| not zero-shot learning. |
|
|
|
|
