--- language: - en license: other size_categories: - 10K 0.3) ``` **Note on joint count:** The source README and paper claim 18 body keypoints. The actual data ships 24: `top_of_head`, `nose`, `neck`, `{right,left}_{shoulder,elbow,wrist,hand,hip,knee,ankle,foot}`, `middle_of_waist`, `{right,left}_{eye,ear}`. **`face_position` → `fo.Keypoints` (per frame)** The face center `(col, row)` from the metadata maps directly to FiftyOne's `(x, y)` — no coordinate swap needed. This field covers only the labeled subject; other people in frame have no face data attached. ### Pose Skeleton ```python dataset.skeletons["pose"] # fo.KeypointSkeleton with 24 labels and 10 bone chains ``` Bone chains: - `top_of_head → nose → neck → middle_of_waist` - `nose → right_eye → right_ear` - `nose → left_eye → left_ear` - `neck → right_shoulder → right_elbow → right_wrist → right_hand` - `neck → left_shoulder → left_elbow → left_wrist → left_hand` - `right_shoulder → right_hip` - `left_shoulder → left_hip` - `right_hip → middle_of_waist → left_hip` - `right_hip → right_knee → right_ankle → right_foot` - `left_hip → left_knee → left_ankle → left_foot` ### Parsing Decisions and Known Data Quality Issues The following issues were discovered by inspecting the actual files (not documented in the source README): | Issue | Scope | Handling | |-------|-------|---------| | Individual pose entries (`poses[i]`) are `None` | All splits, ~10% of frames | Skipped — no `pose` label attached to that frame | | `frame_timestamps` list is `None` | ~11.8% of train cam entries | Synthetic evenly-spaced timestamps generated from `segment_duration_s` | | `poses` key absent entirely from cam entry | 148/34,444 train cam entries | Frame labels skipped for that slice; warned in log | | README claims 18 joints; data has 24 | All splits | All 24 joints stored; `KeypointSkeleton` reflects actual count | ### Action Distribution | Action | Frequency | Count (approx.) | |--------|-----------|-----------------| | `take` | 97.2% | ~20,100 | | `put` | ~2% | ~420 | | `touch` | <1% | ~210 | The heavy `take` imbalance reflects real customer behavior; the distribution was not artificially adjusted. Training models on this data requires weighted sampling or loss weighting for `put` and `touch`. ### Evaluation The competition task (PRAW @ WACV 2027, Track 3: TV-STAL) evaluates predictions with a COCO-style mAP adapted for 2D point localization: - **mAP**: averaged over spatial distance and temporal IoU thresholds - **mAPs**: spatial Average Precision (averaged over lowest temporal IoU threshold) - **mAPt**: temporal Average Precision (averaged over highest spatial distance threshold) - Spatial matching uses a per-video pixels-to-meters factor derived from pose bone lengths **Published baselines (test set):** | Model | Type | mAP | mAPs | mAPt | |-------|------|-----|------|------| | MoViNet-A2 | Conv | 33.5 | 43.8 | **60.9** | | SlowFast-R101 | Conv | 40.2 | 50.4 | 53.2 | | MViT-b | Transf | **41.7** | **55.6** | 58.2 | | ViT-small | Transf | 28.3 | 42.4 | 46.9 | | ViT-base | Transf | 31.1 | 45.7 | 47.0 | | ViT-giant (frozen) | Transf | 38.5 | 50.3 | 58.0 | ## Dataset Creation ### Curation Rationale RetailAction was created to bridge the gap between generic action recognition benchmarks and the specific demands of retail AI. Existing datasets lack the combination of real-world store recordings, multi-view coverage, and fine-grained interaction-point annotations needed for applications like autonomous checkout and shopper behavior analysis. The dataset introduces an interaction-centric annotation paradigm — marking *where* a hand contacts a product rather than drawing a person bounding box — which is more informative for downstream retail applications. ### Source Data #### Data Collection and Processing Data was collected in 10 medium-to-small convenience stores across the United States over multiple years, capturing actual (unscripted) customer visits. Recording devices were 360-degree top-view cameras operating at 30 FPS with a 2880×2880 resolution, ceiling-mounted at an average height of 2.5 meters and positioned so every point in the area of interest is covered by at least two cameras. An automated pipeline processed the raw footage: 1. **2D pose estimation** using a proprietary version of PersonLab [Papandreou et al., ECCV 2018] fine-tuned on 360-degree store footage. 2. **Multi-view 3D pose reconstruction** via tracklet clustering and triangulation. 3. **Interaction interval detection** using a kinematic Graph Convolutional Network [Yan et al., AAAI 2018] operating on 3D poses and shelf positions. 4. **Camera view selection** — the two cameras with the best visibility of the person and their hands are selected per interval. 5. **Frame subsampling** — at most 32 frames are retained per clip using a model that scores frames by hand velocity and acceleration, prioritizing frames with significant movement. 6. **Anonymization** — facial blurring, timestamp anonymization to epoch 1970-01-01T00:00:00, and removal/blurring of store name references. #### Who are the source data producers? Real customers visiting operational convenience stores. All shoppers were given notice of recordings and signed terms of service with Standard AI before data collection. No actors or scripted interactions were used. ### Annotations #### Annotation process Annotation followed a two-step process: 1. **Binary classification and quality labeling** — annotators marked each segment as containing an interaction or not, and flagged data quality issues (bad camera view, low resolution, too few frames, pose errors). A model-in-the-loop strategy iterated three times, reviewing the 10% most disagreeing samples after each round. 2. **Spatio-temporal fine-grained labeling** — annotators precisely marked the temporal start and end of each individual interaction and placed a 2D spatial point on the exact location where the subject's hand contacts the product, for both camera views. In multi-person clips, a red dot on the subject's head in every frame guided annotators to label only the designated individual. Action categories are mutually exclusive per interaction: `take` (picking up an item), `put` (placing an item back), and `touch` (hand contact without taking or placing). #### Who are the annotators? [More Information Needed] #### Personal and Sensitive Information All individuals in the dataset have been anonymized: - Faces are blurred in every video frame - All timestamps are reset to epoch 1970-01-01T00:00:00 - Store name references are removed or blurred - Person identifiers across splits are withheld; shopper identity is not released ## Citation **BibTeX:** ```bibtex @inproceedings{mazzini2025retailaction, title={RetailAction: Dataset for Multi-View Spatio-Temporal Localization of Human-Object Interactions in Retail}, author={Mazzini, Davide and Raimondi, Alberto and Abbate, Bruno and Fischetti, Daniel and Woollard, David M.}, booktitle={ICCV Retail Vision Workshop}, year={2025} } ``` **APA:** Mazzini, D., Raimondi, A., Abbate, B., Fischetti, D., & Woollard, D. M. (2025). RetailAction: Dataset for multi-view spatio-temporal localization of human-object interactions in retail. *ICCV Retail Vision Workshop*. https://doi.org/10.1109/iccvw69036.2025.00250 ## More Information - Source repository: https://huggingface.co/datasets/standard-cognition/RetailAction - Competition: [PRAW @ WACV 2027 — Track 3: TV-STAL](https://www.kaggle.com/competitions/retail-action-challenge-praw-workshop-wacv-2027/overview) - Contact: {davide, bruno, david.woollard}@standard.ai ## Dataset Card Authors Harpreet Sahota ## Dataset Card Contact [More Information Needed]