--- size_categories: - 1M

Open-access skeleton annotations for [HUI360](https://hucebot.github.io/hui360/), a large-scale 360° egocentric dataset for human-robot interaction anticipation in the wild. This repository provides the annotations as **tabular CSV files** (one row per detection), ready for training and evaluation with [HUI360-Baselines](https://github.com/hucebot/HUI360-Baselines/). ## Related resources | Resource | Link | | --- | --- | | Project website | [hucebot.github.io/hui360](https://hucebot.github.io/hui360/) | | Paper (Hugging Face) | [huggingface.co/papers/2608.11051](https://huggingface.co/papers/2608.11051) | | Paper (HAL) | [hal.science/view/index/docid/5609928](https://hal.science/view/index/docid/5609928) | | Videos & JSON annotations (gated) | [rlorlou/HUI360-Videos](https://huggingface.co/datasets/rlorlou/HUI360-Videos) | | SSUP processed videos (gated) | [rlorlou/HUI360-Videos-SSUP](https://huggingface.co/datasets/rlorlou/HUI360-Videos-SSUP) | | Annotation & processing pipeline | [Interact360](https://github.com/RaphaelLorenzo/Interact360/) | | Baselines code | [HUI360-Baselines](https://github.com/hucebot/HUI360-Baselines/) | > **Annotation tool.** Annotations 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. The original per-frame JSON format is available in the `annotations/` folder of [rlorlou/HUI360-Videos](https://huggingface.co/datasets/rlorlou/HUI360-Videos). ## Dataset overview The dataset covers **99 recordings** (70 from INRIA Shelfy, 29 from Cornell SSUP-HRI), subdivided into episodes. Each CSV file corresponds to one *recording*; each row corresponds to one detection of one track. Tracks have a unique ID within a file. Extract all detections for a single track using `unique_track_identifier`: ```python track_data = df[df["unique_track_identifier"] == "2022_09_21_astor_place_landfill_0000_0"] ``` Within each recording, episodes are contiguous in time (no detections were found between episodes, so intervening data was discarded). ## Column reference - `xmin`, `xmax`, `ymin`, `ymax` — bounding box in pixel coordinates. Images are equirectangular (3840×1920); boxes may wrap around the panorama (`xmin > 3840` means `xmin = 4240` is equivalent to `xmin = 400`). - `sapiens_308_[JOINTNAME]_[x,y,score]` — pixel coordinates and confidence for detections using Sapiens with the Goliath 308-keypoint format. - `vitpose_[JOINTNAME]_[x,y,score]` — pixel coordinates and confidence for detections using ViTPose with the COCO-17 format. - `mask_rle` — RLE-encoded binary mask of the person in the image. 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 ``` ## 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} } ```