# COCO-Search18 — Stratified Working Subset (standalone)

Self-contained subset for the GazeAgent human-vs-MLLM visual-exploration study.
Everything needed to run the experiment on this subset is in this folder — no access
to the full COCO-Search18 download or the COCO annotations is required.

## Contents
```
images/TP/<category>/<id>.jpg          141 target-present stimulus images
images/TA/<category>/<id>.jpg          144 target-absent stimulus images
human_fixations/fixations_tp.json      human scanpaths for the TP images (all subjects)
human_fixations/fixations_ta.json      human scanpaths for the TA images (all subjects)
image_ids_tp.json / image_ids_ta.json  frozen id lists + per-image strata (ecc / clutter)
subset_manifest.json                   seed, thresholds, shortfalls — how it was drawn
config.yaml                            the knobs used to build it
```

## Counts
- TP: 141 images, 1410 human scanpaths
- TA: 144 images, 1440 human scanpaths
- 10 human subjects per image.

## Coordinate space (IMPORTANT)
All gaze was recorded on images resized to 1680x1050
(zero-padded, aspect ratio preserved). Fixation `X`/`Y` are pixels in that frame.
Normalize to [0,1] as `x / 1680`, `y / 1050`.

## human_fixations schema (raw COCO-Search18, lossless; one record per image x subject)
`name`, `subject`, `task` (target category), `condition` (present|absent),
`bbox` [x,y,w,h] (TP only — the target), `X`[], `Y`[], `T`[] (fixation x / y / duration
sequences, incl. the initial centre fixation), `length`, `correct`, `RT` (TP) /
`fixOnTarget` (TA), `split`.

## image_ids_* schema (one record per selected image)
`image_id`, `coco_id`, `target_category`, `condition`, `split`, `n_subjects`,
`object_count`, `clutter_bin` (low|high), `stratum`; TP additionally has `bbox`,
`eccentricity`, `ecc_bin`.

## Provenance
- Pool: **validation** split of COCO-Search18 (test-split fixations are
  withheld upstream, so validation is the held-out evaluation set).
- Seed: 0. Clutter = number of COCO-2017 instance annotations per image.
- Stratification: TP = category x eccentricity x clutter; TA = category x clutter;
  thresholds are per-category medians.
- Regenerate: `python scripts/build_subset.py && python scripts/export_subset.py`.

## Cite the source dataset
Yang, Z., Huang, L., Chen, Y., Wei, Z., Ahn, S., Zelinsky, G., Samaras, D., & Hoai, M.
(2020). Predicting Goal-directed Human Attention Using Inverse Reinforcement Learning. CVPR.