README.md

AgentHOI evaluation

Benchmarks, four-metric runners, and per-baseline comparison launchers used for the AgentHOI paper.

Layout

evaluation/
├── README.md                 # this file
├── env.example.sh            # copy -> env.local.sh and fill paths
├── benchmarks/
│   ├── datasets_portable/    # portable test JSONs (relative paths only)
│   └── manifests/            # per-metric manifests (Object-DINO / TVA)
├── scripts/
│   ├── prepare_eval_jsons.{py,sh}
│   ├── prepare_baselines_jsons.sh
│   ├── prepare_benchmark_assets.{py,sh}
│   ├── absolutize_dataset_json.py    # turn portable JSON into absolute-path JSON
│   ├── eval_common.sh                # shared helpers (no internal paths)
│   ├── run_all_metrics.sh            # 4 metrics in one pass
│   ├── run_eval.sh                   # one-liner: bench + method
│   ├── aggregate_results.{py,sh}
│   ├── copy_eval_videos.sh           # stage videos under videos/<bench>/<method>/
│   ├── package_hf_upload.sh          # tarball assets/results for HF dataset upload
│   ├── unpack_hf_download.sh         # inverse of the above
│   └── setup_deps.sh                 # symlink the metric implementations repo
├── baselines/                        # per-method comparison launchers
├── docs/                             # metric notes
├── videos/                           # populated by you / copy_eval_videos.sh
└── results/                          # written by run_all_metrics.sh

Metrics (4 suites)

#	Metric	Conda env (suggested)	Output
1	Object-DINO	dino_obj	obj_dino_result.json
2	VBench	vbench	vbench_results/
3	TVA (VideoAlign / VideoReward)	videoreward	TVA_result.json
4	InternVL	internvl	internvl/summary.json

VBench dimensions: motion_smoothness, dynamic_degree, aesthetic_quality, overall_consistency. InternVL scores object consistency, human consistency, and interaction quality.

Prerequisites

1. Metric implementation repo — VBench / VideoAlign / Object-DINO / InternVL are not bundled (they have heavy weights and large codebases). Point EVAL_DEPS_ROOT at a local clone:

   bash scripts/setup_deps.sh /path/to/your/evaluations_repo
   export EVAL_DEPS_ROOT=/path/to/your/evaluations_repo
   

2. Conda envs (or set PYTHON_* in env.local.sh): one env per metric.

3. Aux weights: VideoReward / VideoAlign checkpoint, InternVL3.5-38B, VBench cache.

4. Generated videos: one .mp4 per image_id in the benchmark JSON, under videos/<benchmark>/<method>/.

Quick start

All paths are relative to release/evaluation/. The repo ships one demo case (benchmarks/datasets_portable/demo_crosstest.json) using the bundled examples/I2V/ reference images, so you can sanity-check the pipeline on a single sample before downloading the full benchmarks.

cd release/evaluation

# 1) Once: link metric code (VBench / VideoAlign / Object-DINO / InternVL)
bash scripts/setup_deps.sh /path/to/your/evaluations_repo

# 2) Optional: set custom PYTHON_* / weight paths
cp env.example.sh env.local.sh
# (edit env.local.sh)

# 3) Run inference for the demo case (from repo root):
#    cd .. && bash run_inference.sh --dataset_path evaluation/benchmarks/datasets_portable/demo_crosstest.json
#    -> writes outputs/demo/<image_id>.mp4
#
# 4) Stage the resulting mp4 under videos/<bench>/<method>/:
#    mkdir -p videos/crosstest/ours
#    cp ../outputs/demo/obj=img_0=human=img_44.mp4 videos/crosstest/ours/

# 5) Run the four metrics:
bash scripts/run_eval.sh crosstest ours
# equivalent: bash baselines/run_ours_crosstest.sh

# 6) Aggregate
bash scripts/aggregate_results.sh
# -> results/eval_summary.xlsx

To use your own video directory:

export VIDEO_PATH=/your/mp4/dir
bash scripts/run_eval.sh crosstest custom

Benchmarks

Three test sets used in the paper. Datasets are NOT bundled (they reference copyrighted source images and are too heavy for git); see benchmarks/datasets_portable/demo_crosstest.json for the schema.

Benchmark	Suggested file	Notes
Cross-test	hoi_crosstest_<date>.json	Main paper comparison (72 clips)
AC	hoi_AC_test_evalset.json	AnchorCrafter-aligned test set
HOMA	hoi_selfcollect_homa1_evalset.json	In-the-wild self-collected set

Use the matching manifests (obj_dino_<bench>.json, tva_<bench>.json) and the corresponding launcher under baselines/.

Baseline comparison workflow

1. Run inference for ours and each baseline on the same benchmark JSON (the video generation step in the main README).
2. Stage each method's mp4 under videos/<benchmark>/<method>/, or set VIDEO_PATH per run.
3. Launch baselines/run_<method>_<benchmark>.sh (or bash scripts/run_eval.sh <benchmark> <method>).
4. bash scripts/aggregate_results.sh to build a comparison Excel.

Compared methods in the paper include VACE, HuMo, HOMA, AnchorCrafter; add a new method by copying any baselines/run_*.sh.

Notes

• Paths in dataset JSON (ref_image_path, ...) should be relative to the evaluation root or absolute on the running machine. Use scripts/absolutize_dataset_json.py to materialize an absolute-path JSON on a new host.
• Large artifacts (VBench weights, full benchmark videos, results/) are not bundled. Keep them under EVAL_DEPS_ROOT or download separately.

Relation to training / inference

Stage	Repo path
Inference (generate videos to evaluate)	../run_inference.sh
Eval benchmark as inference JSON	benchmarks/datasets_portable/demo_crosstest.json (sample)

Evaluation does not load the HOI transformer; it only reads .mp4 files.