gaze2hoi/EgoAIM-Demo

EgoAIM Demo

Portable examples for Gaze2HOI: Gaze-Aware Intention Disambiguation for Egocentric Hand-Object Interaction Forecasting.

Paper · Code · Model checkpoints · Project page

This package contains 11 examples shown on the Gaze2HOI project page: four qualitative comparisons, three failure cases, and four overview examples spanning Establish Contact, Move, Manipulate, and Put Down. It supports both standalone intention inference and the complete intention-to-motion pipeline.

Each example includes one second of observations and a separate one-second reference future at 30 Hz. Inputs contain the original five RGB frames used by the intention model, dual-channel gaze maps, observed 3D hand states, all candidate object tracks, and scene geometry. Model weights and inference code are distributed separately.

Quick start

After installing the inference code, download the examples and checkpoints from the code checkout:

gaze2hoi download \
  --model-repo gaze2hoi/Gaze2HOI \
  --data-repo gaze2hoi/EgoAIM-Demo

The default layout places this directory beside the code checkout. Use --data to select a different data directory. Follow the code repository's MANO setup instructions before running motion inference.

workspace/
├── gaze2hoi-release/
├── egoaim-demo/
└── gaze2hoi-checkpoints/

Use manifest.json to discover examples. The qualitative and failure groups form the default gallery; overview contains the four additional examples from the introductory animation. Every sample records its own draw index and noise profile so that the corresponding reference can be reproduced with the appropriate configuration.

Examples

Draw numbers below are one-based; the machine-readable draw_index is zero-based.

Group	Sample ID	Phase	Draw
Qualitative	P0001_550ea2ac_001680_001739	Establish Contact	3
Qualitative	P0001_f6cc0cc8_000985_001044	Move	1
Qualitative	P0009_8c0caf6c_000215_000274	Move	1
Qualitative	P0001_95eabc0a_000415_000474	Put Down	1
Failure	P0003_ebdc6ff7_001230_001289	Manipulate	2
Failure	P0011_30a5b61d_000545_000604	Establish Contact	2
Failure	P0001_83b755aa_000125_000184	Establish Contact	1
Overview	P0015_3c7f5241_000090_000149	Establish Contact	5
Overview	P0014_9a25ec6a_002215_002274	Move	1
Overview	P0002_59a84a3a_000615_000674	Manipulate	2
Overview	P0001_a9d6c83d_000360_000419	Put Down	5

These examples demonstrate the method and its failure cases. They are not a replacement for the complete EgoAIM benchmark or its aggregate evaluation protocol.

Directory layout

egoaim-demo/
├── README.md
├── LICENSE.md
├── NOTICE.md
├── manifest.json
├── checksums.sha256
├── licenses/
├── shared/
│   ├── intention_schema.json
│   ├── intent_vocab.json
│   ├── object_vocab.json
│   ├── furniture_vocab.json
│   ├── object_references/
│   └── render_objects/
└── samples/<sample_id>/
    ├── sample.json
    ├── inputs/
    │   ├── intention.json
    │   ├── gaze.npz
    │   ├── rgb/
    │   ├── observation.npz
    │   ├── observation.mp4
    │   ├── objects.npz
    │   └── placement.json
    ├── ground_truth/
    │   ├── intention.json
    │   ├── motion.npz
    │   └── scene_motion.npz
    ├── reference_outputs/
    │   ├── intention.json
    │   ├── motion.npz
    │   └── *.mp4
    └── visualization/
        ├── config.json
        └── furniture.npz

All paths in the package are relative. Image paths and schema paths in inputs/intention.json are resolved from the sample directory; shared assets can be referenced with ../../shared/. Array files use NumPy NPZ without pickled objects and can be loaded with allow_pickle=False.

Input contract

Intention inference

inputs/intention.json contains the exact system/user messages, ordered image paths, candidate-object vocabulary, permitted object parts, candidate targets, and gaze-map path. It does not contain an assistant answer.

The first five images are clean RGB observations in chronological order, sampled at offsets 5, 11, 17, 23, and 29 of the 30-frame observation window. Remaining images are candidate-object reference views in the order used by the prompt. The prepared RGB images are 672 x 672; reference images are 448 x 448.

inputs/gaze.npz stores heatmaps with shape [5, 2, 42, 42]. Channels are current and history. These are the original TGSM inputs, including fixation weighting, recency weighting, and scene-aligned gaze projection.

Motion forecasting

inputs/observation.npz contains only observed temporal data:

Field	Shape	Description
state	[30, 66]	Observed hand state; object channels are zero until predicted-intention binding
left_hand_valid, right_hand_valid	[30]	Observed hand availability
mano_betas	[10]	Subject hand-shape coefficients
left_j0, right_j0	[3]	Wrist-root correction terms
scene_present	[33]	Candidate object presence
scene_transl_last, scene_transl_mean	[33, 3]	Object position summaries computed from observations
scene_rot6d_last	[33, 6]	Last observed object orientation
furn_*	Various	Furniture geometry and presence
gaze_*, head_*	30 temporal entries	Observed gaze and head states
anchor_rotation, anchor_origin	[3, 3], [3]	World-to-anchor transform

The 66-dimensional state consists of two 24-dimensional hand blocks and two 9-dimensional object blocks. A hand block contains translation (3), wrist rotation (6), and MANO finger PCA coefficients (15); an object block contains translation (3) and rotation (6). Rotations store the first matrix column followed by the second column and are decoded with Gram-Schmidt orthogonalization.

inputs/objects.npz contains all candidate objects, with state [N, 30, 9], valid [N, 30], and slots, names, and uids [N]. The runtime binds objects after intention inference. The observed contact fallback for an idle hand is stored as contact_left and contact_right in sample.json.

inputs/placement.json contains the calibrated camera intrinsics, observed camera pose, gravity-alignment rotation, and furniture support-plane heights required to lift a predicted image-space placement point into 3D.

Coordinates and timing

Positions are in metres. The anchor frame is gravity-aligned, with +Z up, its origin at the last observed head position, and the horizontal head-forward direction aligned with +X. Hand translations include the MANO wrist-root correction. anchor_rotation maps world vectors to anchor vectors; a point is transformed as R @ (p_world - anchor_origin).

RGB follows the rectified display orientation, rotated 90 degrees clockwise from the native linear camera orientation. Intention placement coordinates use a 0-1000 image grid, with origin at the upper-left corner. The camera calibration in placement.json retains the native linear orientation; the runtime handles this display rotation before unprojection.

Ground truth and archived references

ground_truth/ is provided for comparison. It contains the structured intention and the canonical future hand/object trajectories, including their availability masks. The canonical motion targets preserve EgoAIM's annotated interaction horizon. These files are not inference inputs.

reference_outputs/ contains archived Gaze2HOI predictions associated with the project-page examples. The motion arrays retain the model's numerical output without temporal smoothing or refinement. Reference hand_valid masks preserve the original rendering's hand availability and are used only when visualizing archived references.

For the seven qualitative/failure examples, website_prediction.mp4 is the original project-page prediction clip. The four overview examples include prediction.mp4, rendered from the corresponding archived forecast; the project page adds explanatory staging and overlays. Rendering settings, camera poses, crop boxes, and object appearance are recorded in visualization/config.json.

A fresh model run should write to a separate output directory. Reference arrays and video clips are comparison artifacts and must not be presented as the output of a newly executed model.

Reproducibility

All examples use the same frozen motion checkpoint, 50 DDIM steps, guidance scale 1.5, and eta 0. The VLM uses constrained deterministic decoding. Each sample.json records the original noise profile:

• sample_draw_cpu: a per-sample, per-draw seed derived from the base seed 999; initial noise is generated on the CPU. Used by the qualitative and failure groups.
• sequential_device: one recorded sample seed with sequential draws on the inference device. Used by the overview group.

The profile is part of the configuration: replacing one noise-generation convention with the other changes the forecast even with the same base seed. Exact numerical agreement can also depend on the PyTorch/CUDA versions, device, precision, and batching. Preserve the supplied configuration when comparing against references.

Attribution and terms

These examples are derived from HOT3D, captured with Project Aria, with EgoAIM intention annotations and Gaze2HOI forecasts. The source dataset has component-specific terms; see LICENSE.md and NOTICE.md. MANO model files are not included and must be obtained separately from the official MANO website.

Please cite the Gaze2HOI paper and the upstream HOT3D dataset when using these examples in research. The project page provides project information.