# Evaluation benchmarks

This release ships **schemas and a single demo entry**; the full benchmark
JSONs and reference assets are not bundled. Use the schemas below to materialize
your own copies, or download the official benchmark archive separately.

## Layout

```
benchmarks/
├── README.md
├── datasets/                  # raw dataset JSONs (not bundled, gitignored)
├── datasets_portable/         # portable JSONs with relative ref paths
│   └── demo_crosstest.json    # bundled single-entry example (uses examples/I2V/)
├── manifests/                 # per-metric manifests (Object-DINO / TVA), not bundled
└── assets/                    # per-sample reference images (not bundled)
```

## Dataset JSON schema

Each JSON is a list of dicts, one entry per evaluation case. Required fields:

| Field                      | Description                                                 |
|----------------------------|-------------------------------------------------------------|
| `image_id`                 | unique id (used as `<image_id>.mp4`)                        |
| `caption`                  | text prompt                                                 |
| `ref_image_path`           | human reference image (relative to `evaluation/` if portable) |
| `ref_obj_image_path`       | object reference image                                      |
| `ref_obj_clip_image_path`  | object image for the CLIP encoder; usually same as the above |

Optional fields used by some metrics:

| Field          | Used by   | Description                                |
|----------------|-----------|--------------------------------------------|
| `audio`        | (unused)  | matching speech track (`.wav`)             |
| `dwpose`       | (unused)  | precomputed pose (`.pkl`)                  |
| `fps`          | (unused)  | source video FPS                           |

## Three benchmarks (used in the paper)

| Benchmark      | Suggested filename                    | Samples |
|----------------|---------------------------------------|---------|
| **Cross-test** | `hoi_crosstest_<date>.json`           | 72      |
| **AC**         | `hoi_AC_test_evalset.json`            | 71      |
| **HOMA**       | `hoi_selfcollect_homa1_evalset.json`  | 110     |

## Portable JSONs vs raw JSONs

- `datasets/` — original JSONs with absolute paths into your local data store.
- `datasets_portable/` — same schema but with relative paths under
  `evaluation/assets/<image_id>/`. Use these for release / portable evaluation.
  Convert back and forth with `scripts/absolutize_dataset_json.py` and
  `scripts/prepare_benchmark_assets.{py,sh}`.

## Metric manifests

`manifests/` holds pre-generated manifests for Object-DINO and TVA. Regenerate
from a dataset JSON via:

```bash
bash ../scripts/prepare_eval_jsons.sh          # crosstest
bash ../scripts/prepare_baselines_jsons.sh     # AC + HOMA
```

| Manifest                                            | Used for          |
|-----------------------------------------------------|-------------------|
| `obj_dino_crosstest.json`, `tva_crosstest.json`     | Cross-test        |
| `obj_dino_AC.json`, `tva_AC.json`                   | AC benchmark      |
| `obj_dino_homa.json`, `tva_homa.json`               | HOMA benchmark    |

InternVL is fed the raw dataset JSON directly (no manifest).