Replace Track 1 with 30 selected single-view sequences

README.md

@@ -1,120 +1,12 @@
 ---
-license: cc-by-4.0
-pretty_name: Video to Data (V2D) Challenge Dataset
-tags:
-  - robotics
-  - robot-learning
-  - manipulation
-  - human-object-interaction
-  - motion-capture
-  - video
-  - 4d-reconstruction
-  - reinforcement-learning
-  - egocentric-video
+license:
+  - cc-by-4.0
 ---
 
-# Video to Data (V2D) Challenge Dataset
+# Video-to-Data Challenge
 
-## Dataset Description
-
-The **Video to Data (V2D) Challenge Dataset** is an NVIDIA-developed benchmark for studying the complete path from human demonstration video to physics-grounded robot behavior. It supports three coupled challenge tracks over shared manipulation tasks: **4D human-object interaction reconstruction**, **robotic grounding**, and **end-to-end egocentric transfer**.
-
-- [Challenge website](https://nvidia-isaac.github.io/video_to_data/v2d_challenge/)
-- [Starter toolkit](https://github.com/nvidia-isaac/video_to_data)
-- [Dataset repository](https://huggingface.co/datasets/nvidia/video_to_data_challenge)
-- Contact: [v2d_challenge@nvidia.com](mailto:v2d_challenge@nvidia.com)
-
-## Challenge Tracks
-
-| Track | Input | Goal | Evaluation summary |
-| --- | --- | --- | --- |
-| **Track 1: Reconstruction** | Monocular third-person RGB video | Recover the human, object pose, and object geometry as a metric 4D human-object interaction scene in a consistent world frame | Reconstruction accuracy and physical plausibility relative to the multi-view reference |
-| **Track 2: Robotic Grounding** | Third-person 4D human-object interaction trajectories at different input-noise tiers | Retarget the demonstration and learn an executable policy for the robotic embodiment and simulator | Object-tracking performance at each input tier |
-| **Track 3: Egocentric** | Egocentric human demonstration video | Produce an executable robot policy through either an explicit reconstruct-and-retarget pipeline or an implicit end-to-end method | Final results produced with the official evaluation script |
-
-### Track 1: Reconstruction
-
-Track 1 evaluates monocular 4D human-object interaction reconstruction under challenging conditions including occlusion, bimanual coordination, and long-horizon manipulation.
-
-Participants reconstruct:
-
-- the human body and hands;
-- object pose trajectories;
-- object geometry; and
-- metric scale.
-
-Track 1 is evaluated along two equally weighted axes:
-
-1. **Accuracy**
-   - Chamfer distance to the multi-view human mesh
-   - Chamfer distance to the multi-view object mesh
-2. **Physical plausibility**
-   - Human-joint acceleration error
-   - Object acceleration error
-   - Contact penetration error
-
-### Track 2: Robotic Grounding
-
-Track 2 measures how upstream reconstruction quality affects human-to-robot transfer and downstream policy learning.
-
-The dataset provides three input tiers:
-
-1. **Tier 1 — Clean multi-view capture:** an upper-bound input for upstream reconstruction.
-2. **Tier 2 — Synthetic corruption:** trajectories with jitter, dropout, and contact errors sampled from Track 1 error distributions.
-3. **Tier 3 — Off-the-shelf reconstruction:** trajectories produced by current reconstruction methods.
-
-Each tier is scored separately. Metrics include **AUC**, **SP-SR**, **MP-SR**, and **MPPE**, as defined by the challenge evaluation protocol.
-
-### Track 3: Egocentric
-
-Track 3 evaluates the full pipeline from egocentric human video to robot behavior. The track is method-agnostic: participants may use an explicit reconstruction-and-retargeting pipeline, an end-to-end model, a pretrained vision-language-action model, a world-action model, or a hybrid approach.
-
-The NVIDIA-collected data includes human-object manipulation recordings and associated assets made available for development and evaluation. Depending on the released split, these assets may include:
-
-- egocentric videos;
-- motion-capture trajectories;
-- sequence metadata;
-- textured 3D object meshes; and
-- URDF object descriptions.
-
-Use `eval_e2e.py` from the starter toolkit to package the required reconstructions and recorded policy evaluations.
-
-## Download
-
-Install the Hugging Face Hub client:
-
-```bash
-python -m pip install -U "huggingface_hub"
-```
-
-Download the complete dataset repository while preserving its file structure:
-
-```bash
-hf download nvidia/video_to_data_challenge \
-  --repo-type dataset \
-  --local-dir ./video_to_data_challenge
-```
-
-If authentication is requested, first run:
-
-```bash
-hf auth login
-```
-
-The same operation can be performed from Python:
-
-```python
-from huggingface_hub import snapshot_download
-
-snapshot_download(
-    repo_id="nvidia/video_to_data_challenge",
-    repo_type="dataset",
-    local_dir="./video_to_data_challenge",
-)
-```
-
-Large assets are stored using Hugging Face's large-file infrastructure. Make sure sufficient disk space is available before downloading the complete repository.
-
-## Support
-
-For challenge or dataset questions, contact [v2d_challenge@nvidia.com](mailto:v2d_challenge@nvidia.com). For software issues, use the issue tracker in the [Video to Data repository](https://github.com/nvidia-isaac/video_to_data/issues).
+- [Track 2](track_2/README.md): the dataset and all accompanying assets are licensed
+  under [CC BY-4.0](https://creativecommons.org/licenses/by/4.0/).
+- [Track 3](track_3/README.md): the dataset and all accompanying assets are licensed
+  under [CC BY-4.0](https://creativecommons.org/licenses/by/4.0/). See the
+  [Track 3 dataset card](track_3/v2d-challenge-track3.md).

track_2/README.md

@@ -1,12 +1,13 @@
 # Track 2: trajectory variants
 
-Track 2 contains the same 30 human-object interaction episodes in two challenge tiers. Both tiers use the same episode order, 22,990-frame timeline, tasks, exocentric RGB observation, fitted ground planes, and G1 support surfaces.
+Track 2 contains the same 30 human-object interaction episodes in three parallel trajectory variants. All variants use the same episode order, 22,990-frame timeline, tasks, and four synchronized RGB views.
 
-| Tier | Directory | Human trajectory | Object trajectory and mesh |
+| Variant | Directory | Human trajectory | Object trajectory and mesh |
 |---|---|---|---|
-| Tier 1: multiview caption | [`tier_1_multiview_caption/`](tier_1_multiview_caption/) | Original SOMA-X reconstruction | Original pose and aligned mesh |
-| Tier 2: synthetic noise | [`tier_2_synthetic_noise/`](tier_2_synthetic_noise/) | Synthetically perturbed human trajectory | Synthetically perturbed object trajectory; original mesh |
+| Baseline | [`baseline/`](baseline/) | Original four-view SOMA-X reconstruction | Original four-view pose and aligned mesh |
+| Synthetic random walk | [`synthetic_random_walk/`](synthetic_random_walk/) | Baseline with deterministic root SE(3) random walk | Baseline with an independent deterministic object SE(3) random walk; original mesh |
+| CARI4D reconstruction | [`cari4d_reconstruction/`](cari4d_reconstruction/) | Monocular CARI4D reconstruction converted to SOMA-X and rigidly aligned from the first-frame human mesh | The coordinate-equivalent alignment in the source OpenCV world, with poses expressed in the baseline canonical object mesh frame |
 
-Each directory is a complete LeRobot v2.1 dataset with its own `README.md`, `meta/`, `data/`, `videos/`, and `mesh/`. Select one tier directory as the dataset root. Do not concatenate tiers: episode and global frame indices intentionally overlap so results can be compared frame by frame.
+Each directory is a complete LeRobot v2.1 dataset with its own `README.md`, `meta/`, `data/`, `videos/`, and `mesh/`. Select one variant directory as the dataset root. Do not concatenate variants: episode and global frame indices intentionally overlap so results can be compared frame by frame.
 
-The `observation.images.exo_camera` video stream is byte-identical across tiers and is the monocular input participants should use for reconstruction. “Multiview caption” is the challenge tier name; this staged participant release exposes only the selected exocentric stream. Reconstruction results are not published; participants generate them with their own method before retargeting and grounding.
+The four video streams are byte-identical across variants. The synthetic variant uses deterministic seed-42 cumulative bounds of 16 mm / 8 degrees for the human root and 30 mm / 30 degrees for the object. These produce 3.815 cm CD-h and 6.045 cm CD-o, calibrated to half of CARI4D's BEHAVE paper errors. The CARI4D variant is reconstructed from the front-left video and masks; the other three streams are retained as synchronized observations. Its first-frame alignment removes global coordinate offsets with one human-derived transform per episode. Because SOMA-X human data and object poses use MHR-native and OpenCV world coordinates respectively, the object transform includes their fixed Y/Z basis conversion. It does not independently fit the object trajectory. CARI4D-style CD-h and CD-o distributions are included in both variant provenance files.

track_2/{tier_2_synthetic_noise → baseline}/README.md

@@ -1,8 +1,10 @@
-# Track 2 — Tier 2: synthetic noise
+# Track 2 — baseline
 
-This tier provides synthetically perturbed human and object trajectories while
-preserving identity parameters, visibility, episode timing, videos, and meshes.
-Object poses remain **`world_T_object` in the source OpenCV world frame**.
+The original Track 2 human-object interaction sequences reconstructed with SOMA-X from four
+frame-synchronized RGB views. Every episode includes a full-body pose and one tracked
+object with a textured 3D mesh. Object poses are expressed as **`world_T_object` in the
+source OpenCV world frame**, so placing the mesh at a logged pose recovers its
+reconstructed location.
 
 | | |
 |---|---|
@@ -10,23 +12,23 @@ Object poses remain **`world_T_object` in the source OpenCV world frame**.
 | Frames | 22,990 (440–1,003 per episode, mean 766) |
 | Rate | 30 fps nominal |
 | Tasks | 30 |
-| Cameras | 1 exocentric stream, 30 clips |
+| Cameras | 4 streams, 120 clips |
 | Objects | 29, each with a textured mesh |
 
 ## Layout
 
 ```
-track_2/tier_2_synthetic_noise/
+track_2/baseline/
 ├── data/chunk-000/episode_0000NN.parquet    human + object pose, one row per frame
-├── videos/chunk-000/observation.images.exo_camera/episode_0000NN.mp4
+├── videos/chunk-000/observation.images.<camera>/episode_0000NN.mp4
 ├── mesh/<object>/<object>.glb               textured mesh
-├── support_surface/<sequence>_support.usda  static G1 support geometry
 └── meta/
     ├── info.json                            schema, counts, tasks, objects
     ├── episodes_metadata.jsonl              episode → sequence, task, object, mesh
     ├── episodes.jsonl                       episode lengths and tasks
     ├── episodes_stats.jsonl                 per-episode feature statistics
     └── tasks.jsonl                          task descriptions
+
 ```
 
 ## Pose data
@@ -68,11 +70,14 @@ An object's name **is** its mesh folder: `white_desk` →
 
 | Stream | Resolution | Kind |
 |---|---|---|
-| `exo_camera` | 1536×1152 | colour, static third-person |
+| `back_stereo_camera_left` | 1536×1152 | colour, static third-person |
+| `front_stereo_camera_left` | 1536×1152 | colour, static third-person |
+| `left_stereo_camera_left` | 1536×1152 | colour, static third-person |
+| `right_stereo_camera_left` | 1536×1152 | colour, static third-person |
 
-Video is **frame-exact with the parquets** — row *i* corresponds to frame *i* in the
-clip, verified across all 30 episodes. Each file is a byte-identical copy of its selected
-source H.264 clip; no concatenation or re-encoding was performed.
+Video is **frame-exact with the parquets** — row *i* corresponds to frame *i* in every
+clip, verified across all 120. Files are byte-identical copies of the
+source H.264 videos; no concatenation or re-encoding was performed.
 
 The supplied export has **no camera intrinsics or extrinsics**, so meshes cannot be
 projected into the images from this dataset alone. The source calibration sequence
@@ -87,11 +92,3 @@ as `observation.object.pose`. SOMA-to-G1 retargeting transforms this fitted
 plane through its first-frame anchor. Do not replace it with a horizontal
 world-plane estimate when reproducing a retargeted training reference.
 
-## Support surface
-
-Every episode declares `support_surface/<sequence_id>_support.usda` in
-`meta/episodes_metadata.jsonl`. These static collision stages use metres and Z-up in
-the released G1 retargeted world. Place the declared file at
-`<motion-root>/whole_body/reconstructed_stage/<sequence_id>_support.usda` alongside
-the G1 retargeted motion dataset before grounding or training. Ground-only episodes
-have a valid empty stage at the same path.