Upload README.md with huggingface_hub

README.md

@@ -1,101 +1,349 @@
----
-task_categories:
-- video-classification
-tags:
-- Human
-- ComputerVision
-- Motion
-- Keypoints
-size_categories:
-- 1M<n<10M
----
-
-# HUI360 : A 360° Egocentric Dataset and Baselines for Human-Robot Interaction Anticipation
-
-A dataset to predict human-robot Interaction. Under review.
-
-Each file corresponds to a *recording* and each line corresponds to one detection of one track. 
-In each recording there are several *episodes* with contiguous time (no detection were found in-between episodes so data was discarded).
-
-Tracks have a unique ID (among a file) so you may extract all the detections for a single track using `unique_track_identifier` : 
+# HUI360-Videos
+
+**HUI360: A 360° Egocentric Dataset and Baselines for Human-Robot Interaction Anticipation** (IEEE FG 2026)
+
+<p align="center">
+  <a href="https://hucebot.github.io/hui360/"><img src="https://img.shields.io/badge/website-hui360-green" alt="Website"></a>
+  <a href="https://hal.science/view/index/docid/5609928"><img src="https://img.shields.io/badge/paper-HAL%20Science-blue" alt="Paper"></a>
+  <a href="https://doi.org/10.1109/FG67764.2026.11556969"><img src="https://img.shields.io/badge/IEEE-FG%202026-orange" alt="IEEE FG 2026"></a>
+  <a href="https://github.com/RaphaelLorenzo/Interact360/"><img src="https://img.shields.io/badge/annotation%20tool-Interact360-black?logo=github" alt="Annotation tool"></a>
+  <a href="https://github.com/hucebot/HUI360-Baselines/"><img src="https://img.shields.io/badge/baselines-GitHub-black?logo=github" alt="Baselines"></a>
+  <a href="https://huggingface.co/datasets/rlorlou/HUI360"><img src="https://img.shields.io/badge/annotations-HuggingFace-blueviolet" alt="Annotations"></a>
+  <a href="https://huggingface.co/datasets/rlorlou/HUI360-Videos"><img src="https://img.shields.io/badge/videos-HuggingFace-yellow" alt="Videos"></a>
+</p>
+
+Video data for [HUI360](https://hucebot.github.io/hui360/), a large-scale 360° egocentric dataset for human-robot interaction anticipation in the wild.
+
+## Related resources
+
+| Resource | Link |
+| --- | --- |
+| Project website | [hucebot.github.io/hui360](https://hucebot.github.io/hui360/) |
+| Paper (HAL) | [hal.science/view/index/docid/5609928](https://hal.science/view/index/docid/5609928) |
+| Paper (IEEE) | [doi.org/10.1109/FG67764.2026.11556969](https://doi.org/10.1109/FG67764.2026.11556969) |
+| Annotation & processing pipeline | [Interact360](https://github.com/RaphaelLorenzo/Interact360/) |
+| Baselines code | [HUI360-Baselines](https://github.com/hucebot/HUI360-Baselines/) |
+| Skeleton annotations (open) | [rlorlou/HUI360](https://huggingface.co/datasets/rlorlou/HUI360) |
+| SSUP processed videos (gated) | [rlorlou/HUI360-Videos-SSUP](https://huggingface.co/datasets/rlorlou/HUI360-Videos-SSUP) |
+| SSUP original dataset | [IRL-CT/SSUP-HRI](https://github.com/IRL-CT/SSUP-HRI) |
+
+> **Annotation tool.** Annotations in this repository were produced with the [Interact360](https://github.com/RaphaelLorenzo/Interact360/) pipeline (automatic detection, tracking, pose estimation, and interaction labelling) and manually refined with its `visualize.py` GUI. Install Interact360 with the `[gui]` extra to inspect or correct annotations. Fields such as `validity` and `engagement_level` may reflect these manual corrections.
+
+## Dataset sources
+
+This dataset is composed of two sources whose access rules are different:
+
+- Recordings made with the Shelfy Robot at INRIA Nancy in July and October 2025
+- Recordings made by the Interaction Research Lab team of Cornell University in New York in October 2022 and July 2023, reprocessed to provide a unified formatting ([SSUP](https://github.com/IRL-CT/SSUP-HRI) — F. Bu, W. Ju, 2024)
+
+The annotations are freely available and contain the 2D human pose data used to create HUI360 on [Hugging Face](https://huggingface.co/datasets/rlorlou/HUI360).
+
+## Structure and content
+
+### Access rules
+
+The full dataset is gated on Hugging Face and access can only be granted under the following terms:
+
+- Images recorded by the Shelfy Robot at INRIA are available on request for research purposes only. Please submit the form XXXXX.pdf; for non-EU-based institutions you will have to submit a request to your IRB, Ethics Committee, or equivalent.
+- Images recorded by Cornell's IRL team as part of SSUP are subject to their own agreement (available at [IRL-CT/SSUP-HRI](https://github.com/IRL-CT/SSUP-HRI)). As per clause 5, the data cannot be redistributed to users who have not been approved for access to the original dataset. Either request access to the original dataset and split it as specified in (1) below, or — if granted access by the SSUP authors — request access to our processed videos at [rlorlou/HUI360-Videos-SSUP](https://huggingface.co/datasets/rlorlou/HUI360-Videos-SSUP) after sending us proof of access.
+
+### Note
+The dataset is structured in 99 recordings (29 from SSUP Dataset and 70 from Shelfy), each recordings is subdivided in episodes.
+
+(1) In SSUP (XXX_astor_place_XXX or XXX_albee_square_XXX recordings) data extraction the episodes are a simple subdvidision of the original videos every 5 minutes (except for `2022_09_21_astor_place_recycle` that was accidentally divided in 2 minutes sequences)
+(2) In Shelfy recordings (rosbag2_XXX recordings) the episodes have a variable length : they start 10 seconds before someone is detected and end after 10 second without detection (detections were performed with another camera that was only front facing, so people in the back or on the sides are not considered for the episode splitting) 
+
+### Overall structure (after merging)
+- Unzip (if necessary) and merge `video_full_shelfy` and `video_full_ssup` directories into `video_full`
+- Unzip (if necessary) and merge `video_mini_shelfy` and `video_mini_ssup` directories into `video_mini`
+
+The structure is separated for easier downloading :
+```
+HUI360-Video/
+├── README.md
+├── annotations
+│   ├── 2022_09_21_astor_place_landfill
+│   ...
+│   └── rosbag2_2025_10_20-16_02_46
+├── dataset
+│   ├── 2022_09_21_astor_place_landfill
+│   ...
+│   └── rosbag2_2025_10_20-16_02_46
+├── video_full
+│   ├── 2022_09_21_astor_place_landfill
+│   ...
+│   └── rosbag2_2025_10_20-16_02_46
+└── video_mini
+    ├── 2022_09_21_astor_place_landfill
+    ...
+    └── rosbag2_2025_10_20-16_02_46
+```
+
+In details : 
+
 ```
-track_data = df[df["unique_track_identifier"] == "2022_09_21_astor_place_landfill_0000_0"]
+HUI360-Video/
+├── README.md
+├── annotations
+│   ├── 2022_09_21_astor_place_landfill
+│   │   └── episodes
+│   │       ├── 0000
+│   │       │   ├── annotations_360.json
+│   │       │   ├── images_360_filenames.txt
+│   │       │   └── interaction_zone_tracking_360.json
+|   |       ...  
+│   │       └── 0012
+│   │           ├── ...
+│   │           └── ...
+│   ...
+│   └── rosbag2_2025_10_20-16_02_46
+│       └── episodes
+│           ├── 0000
+│           │   ├── annotations_360.json
+│           │   └── images_360_filenames.txt
+|           ...  
+│           └── 0034
+│               ├── ...
+│               └── ...
+├── dataset
+│   ├── 2022_09_21_astor_place_landfill
+│   │   ├── episodes
+│   │   │   ├── 0000
+│   │   │   │   └── images_360
+│   │   │   │       ├── ...
+│   │   │   │       ├── ...
+│   │   │   │       ├── ...
+│   │   │   │       └── ...
+|   |   |   ...  
+│   │   │   └── 0012
+│   │   │       └── ...
+|   │   ├── episodes_flag.json
+│   │   └── episodes.json
+│   ...
+│   └── rosbag2_2025_10_20-16_02_46
+│       ├── episodes
+│       ...
+|       ├── episodes_flag.json
+│       └── episodes.json
+├── video_full
+│   ├── 2022_09_21_astor_place_landfill
+│   │   └── episodes
+│   │       ├── 0000
+│   │       │   └── images_360_full.mp4
+|   |       ...  
+│   │       └── 0012
+│   │           ├── ...
+│   │           └── ...
+│   ...
+│   └── rosbag2_2025_10_20-16_02_46
+│       └── episodes
+│           ├── ...
+│           └── ...
+└── video_mini
+    ├── 2022_09_21_astor_place_landfill
+    │   └── episodes
+    │       ├── 0000
+    │       │   └── images_360_mini.mp4
+    |       ...  
+    │       └── 0012
+    │           ├── ...
+    │           └── ...
+    ...
+    └── rosbag2_2025_10_20-16_02_46
+        └── episodes
+            ├── ...
+            └── ...
 ```
 
-- For each detection columns `xmin, xmax, ymin, ymax` indicates the bounding box in pixel coordinates (note that the images are equirectangular with size 3840x1920 but bounding boxes may appear with xmin > 3840 which indicates a wrapping around : xmin = 4240 is equivalent to xmin = 400)
-- For each detection columns `sapiens_308_[JOINTNAME]_[x,y,score]` contains the pixel coordinates and confidence for detections using Sapiens with Goliath 308 keypoints format
-- For each detection columns `vitpose_[JOINTNAME]_[x,y,score]` contains the pixel coordinates and confidence for detections using VitPose with COCO 17 keypoints format
-- The `mask_rle` column contains the RLE encoded binary mask of the person in the image. RLE encoding / decoding functions :
-
-```python
-import torch
-
-def encode_RLE(mask):
-    """
-    Encode a mask into a RLE.
-    Args:
-        mask: torch.bool [H, W]
-    Returns:
-        runs: torch.tensor [N] - run lengths
-    """
-    flat = mask.flatten() # [H*W]
-    
-    if flat.numel() == 0:
-        return torch.tensor([], device=flat.device), False
-    
-    starts_with_true = flat[0].item()
-    
-    # Find transitions between True/False
-    # Add dummy values at start and end to handle boundaries
-    padded = torch.cat([torch.tensor([not flat[0]], device=flat.device), flat, torch.tensor([not flat[-1]], device=flat.device)])
-    
-    # Find where values change
-    transitions = torch.nonzero(padded[1:] != padded[:-1], as_tuple=False).flatten()
-    
-    # Calculate run lengths
-    runs = torch.diff(transitions)
-    
-    # append a 1 if starts_with_true else a 0 so that we don't have to return starts_with_true
-    if starts_with_true:
-        runs = torch.cat([torch.tensor([1], device=runs.device), runs])
-    else:
-        runs = torch.cat([torch.tensor([0], device=runs.device), runs])
-        
-    return runs
-
-def decode_RLE(runs, shape):
-    """
-    Decode a RLE into a mask.
-    Args:
-        runs: torch.tensor [N] - run lengths
-        shape: tuple - shape to reshape result to
-    Returns:
-        mask: torch.bool [H, W]
-    """
-    
-    start_with_true = runs[0].item()
-    runs = runs[1:]
-    
-    if runs.numel() == 0:
-        return torch.zeros(shape, dtype=torch.bool, device=runs.device)
-    
-    # Create alternating pattern: start_with_true determines first value
-    start_val = 1 if start_with_true else 0
-    vals = (torch.arange(runs.numel(), device=runs.device) + start_val) % 2
-    
-    # Expand runs into full sequence
-    expanded = torch.repeat_interleave(vals, runs).bool()
-    
-    # Reshape to target shape
-    total_elements = shape[0] * shape[1] if len(shape) == 2 else shape[0]
-    if expanded.numel() != total_elements:
-        # Pad or truncate if needed
-        if expanded.numel() < total_elements:
-            padding = torch.zeros(total_elements - expanded.numel(), dtype=torch.bool, device=runs.device)
-            expanded = torch.cat([expanded, padding])
-        else:
-            expanded = expanded[:total_elements]
-    
-    mask_dec = expanded.view(shape)
-    return mask_dec
-``
+### Recording level annotation files
+The files `episodes.json` and `episodes_flag.json` contains informations related to each episode in the recording, the first one is just for the name and timestamp and was created at the extraction phase, it is redundant with the second one that contains specific flags for each episode, anything other than `valid,checked` should not be used. Notable flags are `nov` (episode discarded because of abnormal behaviors attributed to novelty effect), and `GDPR` episode removed because someone requested their data to be removed in compliance with GDPR policy.
+
+- `episodes.json`
+```
+[
+    {
+        "session": "2022_09_21_astor_place_landfill",
+        "episode": "0000",
+        "start_time": 1763542047.679756,
+        "end_time": 1763542064.8130896
+    },
+    {
+        "session": "2022_09_21_astor_place_landfill",
+        "episode": "0001",
+        "start_time": ...,
+        "end_time": ...
+    },
+    ...
+]
+```
+
+- `episodes_flag.json`
+```
+{
+    "0000": {
+        "start_time": 1756981067.0,
+        "end_time": 1756981366.9999995,
+        "start_time_formatted": "2025-09-04 12:17:47",
+        "end_time_formatted": "2025-09-04 12:22:47",
+        "flag": "valid,checked"
+    },
+    "0001": {
+        "start_time": 1756981367.0,
+        "end_time": 1756981666.9999995,
+        "start_time_formatted": "2025-09-04 12:22:47",
+        "end_time_formatted": "2025-09-04 12:27:47",
+        "flag": "valid,checked"
+    },
+    ...
+}
+```
+
+### Episode level annotation and data files
+#### Data
+- `images_360` are empty directories that should normally contain the raw and high resolution panoramic images as .jpeg, they are not kept to save space, please refer to the last section for informations on how to get them (either from the original full quality data or extract them from the videos)
+- `images_360_full.mp4` are compressed videos at full resolution (3840*1920) but whose quality is slightly lower than original JPEGs
+- `images_360_mini.mp4` are compressed videos at reduced resolution (960*480) for very fast downlooading / loading
+
+All the data were recorded or subsampled at 15 fps
+
+#### Annotations
+
+- `interaction_zone_tracking_360.json` are used for SSUP where the "interaction zone" (the trash can) is mobile, the structure for the detections and masks is the same as `annotations_360.json` but with only one detection per image
+- `annotations_360.json` is a list of dictionnaries (one per frame, ordered as in the videos) that contains metadata about the frame and a list of detections (with associated `track_id`) in the frame, these detections are themselves structured a individual dictionnaries, with all the data regarding the track at the given frame (keypoints, bounding box, segmentation mask encoded as RLE). 
+
+    In each detection you will find the following keys :
+
+    - `"track_id"` is assigned to each person tracked along the episode.
+    - The bounding box coordinates are expressed as `xmin`, `ymin`, `xmax` and `ymax` in pixels. As the 360 image is wrapped, some may exceed the image size (but should not be negative) : pixel `(4283,783)` corresponds to pixel `(443,783)`.
+    - `"validity"` flag is assigned (default is `valid`) but it might be corrected in the [Interact360](https://github.com/RaphaelLorenzo/Interact360/) manual correction tool (`visualize.py`)
+    - `"mask_size"` indicates the total number of pixels in the mask
+    - `"mask_rle"` is the RLE encoding of the mask as a string of comma separated values, it can be decoded using `decode_RLE` from `misc.rle_tools.py` (using torch tensor format) : `mask = decode_RLE(torch.from_numpy(np.array(det["mask_rle"].split(",")).astype(np.int32)).to(DEVICE), shape = (width, height))`
+    - `"engagement_level"` is automatically determined by the interaction detection process and can be corrected from the [Interact360](https://github.com/RaphaelLorenzo/Interact360/) manual correction tool (`visualize.py`). `0` corresponds to no engagement, `3` corresponds to interaction (currently happening).
+    - `"intersection_pixels"` is the number of pixels intersecting between the person mask and the interesection zone mask (either fixed or mobile). NB : in older versions this was not exported so `"intersection_pixels" : -1` indicates that this information was not saved.
+    - `"scores_sapiens_308_03b"` or `"scores_sapiens_308_06b"` is a list of float with lenght 308 that are the confidence scores associated with each keypoint of `"keypoints_sapiens_308_03b"` 
+    - `"scores_vitpose"` is a list of float with lenght 17 that are the confidence scores associated with each keypoint of `"keypoints_vitpose"` 
+    - `"keypoints_sapiens_308_03b"` or `"keypoints_sapiens_308_06b"` are the keypoints coordinates (in pixels) with the Goliath format of Sapiens, some were extracted using the 03b pose model and others with 06b pose model. 
+    - `"keypoints_vitpose"` are the keypoints coordinates (in pixels) with the COCO-17 format of ViTPose.
+    The keypoints are expressed as lists of length `308*2` or `17*2` (alternating `x` and `y` coordinates), **note that they are expressed as the pixel position of the joint in the image, so when the box is wrapped around they may end up not being in the box numerically, or showing two parts of the body at the two ends of the image**. 
+    If a `keypoint_XXX` key is not present it means that these keypoints were not extracted, probably because the track was already labelled as invalid or the box is too small.
+
+Overall structure of `annotations_360.json`
+```
+[
+  {
+    "image_file": "1970_01_01_02_00_00_000000.jpg",
+    "image_shape": [
+        1920,
+        3840,
+        3
+    ]
+    "detections": [
+      {
+        "track_id": 0,
+        "xmin": 4283,
+        "ymin": 783,
+        "xmax": 4451,
+        "ymax": 1220,
+        "validity": "valid",
+        "mask_size": 44302,
+        "mask_rle": "0,3007221,17,3820,...,13,2687454",
+        "engagement_level": 3,
+        "intersection_pixels": 18545,
+        "keypoints_vitpose": [
+            481,
+            812,
+            ...
+        ],
+        "scores_vitpose": [
+            0.53,
+            ...
+        ],
+        "keypoints_sapiens_308_03b": [
+            498,
+            816,
+            ...
+        ],
+        "scores_sapiens_308_03b": [
+            0.45,
+            ...
+        ]
+      },
+      {
+        "track_id": 4,
+        "xmin": 1253,
+        "ymin": 721,
+        "xmax": 1611,
+        "ymax": 1520,
+        "validity": "invalid",
+        "mask_size": 145304,
+        "mask_rle": "0,8494527,21,4244,...,14,1425546",
+        "engagement_level": 0,
+        "intersection_pixels": 0,
+        "keypoints_vitpose": [
+            1481,
+            812,
+            ...
+        ],
+        "scores_vitpose": [
+            0.43,
+            ...
+        ],
+        "keypoints_sapiens_308_06b": [
+            1498,
+            816,
+            ...
+        ],
+        "scores_sapiens_308_06b": [
+            0.64,
+            ...
+        ]
+      },
+    ]
+  },
+  {
+    "image_file": "1970_01_01_02_00_00_066667.jpg",
+    "image_shape": [
+        1920,
+        3840,
+        3
+    ]
+    "detections": [
+      ...
+    ]
+  },
+  ...
+]
+```
+
+## Citation
+
+If you use this dataset, please cite:
+
+```bibtex
+@INPROCEEDINGS{11556969,
+  author={Lorenzo-Louis, Raphael and Amadio, Fabio and Luvison, Bertrand and Ivaldi, Serena},
+  booktitle={2026 IEEE 20th International Conference on Automatic Face and Gesture Recognition (FG)},
+  title={HUI360 : A 360° Egocentric Dataset and Baselines for Human-Robot Interaction Anticipation},
+  year={2026},
+  volume={},
+  number={},
+  pages={1-9},
+  doi={10.1109/FG67764.2026.11556969}
+}
+```
+
+If you use the SSUP recordings included in this dataset, please also cite the original SSUP-HRI dataset:
+
+```bibtex
+@inproceedings{BuJ24a,
+  author={Bu, Fanjun and Ju, Wendy},
+  title={SSUP-HRI: Social Signaling in Urban Public Human-Robot Interaction Dataset},
+  booktitle={Social Robotics -- 16th International Conference, ICSR + AI 2024, Odense, Denmark, October 23--26, 2024, Proceedings, Part III},
+  series={Lecture Notes in Computer Science},
+  volume={15563},
+  pages={479--487},
+  publisher={Springer},
+  year={2024},
+  doi={10.1007/978-981-96-3525-2_40}
+}
+```