# AgentHOI: minor-revision evaluation materials Download [agenthoi_minor_revision.zip](agenthoi_minor_revision.zip) and extract it to obtain the `minor_revision/` directory. This release accompanies the additional evaluations in the TOG minor revision of **AgentHOI: Multi-Agent Reasoning for Human-Object-Interaction Video Generation via Implicit Representation Alignment**. ## Contents | Directory | Contents | Paper reference | |---|---|---| | `plan_evaluation/` | Plans under single-agent, w/o clipping refiner and full settings; blinded Qwen3-VL annotations; metric outputs and judge scripts | Table 3; supplementary D | | `agent_ablation/` | Keyboard, cream tube, razor and mask inputs, stage outputs and generated videos for full, w/o Perception and w/o Interaction | Fig. 9–10; supplementary Fig. 2–3 | | `motion_text_matching/` | Candidate descriptions, per-case matching results, aggregate results and example RGB/motion frames | Table 4; supplementary Fig. 4 | | `paper_figures/`, `paper_tables/` | Final figure PDFs and table sources | Same numbering as the revised paper | The fifth and sixth ablation examples from the initial experiment export are excluded. Figures use the final captions and layouts, including the cream-tube hand-transfer example. `M_motion` is the training-only auxiliary motion model. ## Plan evaluation The judge is **Qwen3-VL-32B-Instruct**, independent of the **InternVL3.5-38B** planner family. It receives the human reference, object reference, input caption and timeline without the setting label, with greedy decoding. Each timeline has one-second segments. The reported mechanism metrics are object grounding, explicit acting-hand assignment over contact segments, plan-level hand-state consistency, and the fraction of segments containing more than two motions. Metrics are computed per plan and then averaged. Plans without contact are excluded from the acting-hand metric. Whole-plan body-contact staging and one-step attachment are assessed in a separate pass (`results/contact/`); the preliminary per-segment body-contact labels in `results/mechanism/` are not the paper's body-contact audit. The decomposition improves object grounding and hand specification. The clipping refiner turns abrupt body attachment into staged contact. Some refined phrases leave the acting hand implicit; the video ablation in the paper shows unchanged InternVL-H and improved TVA and InternVL-inter relative to w/o clipping refiner. ### Inspect or reproduce the recorded metrics From the extracted `minor_revision` directory: ```bash python scripts/summarize_release.py python scripts/verify_release.py ``` `summarize_release.py` recomputes the paper's plan metrics directly from the released judge annotations and the motion matching accuracy from per-case rows. The original aggregate JSON files also retain additional diagnostic metrics. ### Rerun the plan judge Download the reference images already provided in this dataset repository: `benchmarks/assets//ref_human.png` and `ref_object.png`. Download the local Qwen3-VL-32B-Instruct checkpoint and install a Transformers version that provides `Qwen3VLForConditionalGeneration`, PyTorch and Pillow. The original experiment used multiple GPUs; the release supports configurable GPU selection. ```bash export AGENTHOI_EVAL_ROOT=/absolute/path/to/agenthoi-eval export PLAN_JUDGE_MODEL=/absolute/path/to/Qwen3-VL-32B-Instruct export PLAN_JUDGE_GPUS=0,1,2,3,4,5,6,7 export PLAN_OUTPUT_DIR=/absolute/path/to/new_judge_results python plan_evaluation/scripts/run_mechanism_qwen.py python plan_evaluation/scripts/mechanism_metrics.py python plan_evaluation/scripts/run_contact_qwen.py ``` `plans.jsonl` image paths are relative to `AGENTHOI_EVAL_ROOT`. Original variant strings are preserved (`Single Agent (prior)`, `w/o Refiner (prior)`, `Full (prior)`); `(prior)` is an export label. The judge never sees that label. Rerunning inference requires the GPU/model dependencies above. Only data aggregation and release integrity checks were executed during packaging. ## Agent ablations and intermediate outputs `manifest.json` maps the four cases to reference images and captions. `stage_traces.json` and `plans/matched_cuda_raw/` preserve the structured outputs, including original agent prompts. `videos/` contains the corresponding full-control, no-perception and no-interaction generations. The keyboard and mask cases illustrate incorrect object categories or unrequested wearing after removing an agent. Fig. 10 uses the cream tube: the refiner inserts alignment with the receiving hand before transfer. The cream-tube video variants show smaller differences than the failure cases. ## Motion–text matching The test pairs each ground-truth HOMA motion sequence with its correct interaction description and three incorrect actions on the same object. The incorrect actions use the same templates: throwing/catching, putting on the floor/kicking, and lifting overhead/swinging. All four candidates receive the same Gaussian noise at noise level 0.5 (timestep 500); the candidate with the lowest flow-matching loss is selected. `M_motion` has 84% top-1 accuracy, the base Wan2.1-T2V-1.3B has 14%, and chance is 25%. The model is used for feature alignment during training and removed at inference. `results/all_cases_comparison.csv` stores measured per-case ranks, losses and top-1 decisions; `data/candidates.json` retains descriptions and encoded-input hashes. Fields such as `input` and `source_video` preserve the original experiment's provenance and are not portable file locations. Example RGB and motion frames are bundled. Full encoded motion latents, the trained `M_motion` checkpoint and its custom inference implementation are not included in the supplied materials, so this release supports reaggregation of the recorded matching results rather than a standalone model-inference rerun. ## Integrity and provenance `SHA256SUMS` records every released file except itself. `PROVENANCE.json` records the source-archive hashes and packaging transformations. The existing dataset's benchmark images and GT videos remain the source inputs.