RetailAction Challenge: PRAW Workshop @ WACV 2027
Welcome to the RetailAction challenge, a competition part of the Physical Retail AI Workshop (PRAW) taking place at the WACV 2027 conference. Standard AI provides the dataset and challenge for this event.
The goal is to develop AI models that can accurately detect and localize shopper-shelf interactions in top-view retail videos. Your model must identify when actions occur (temporal localization) and where they happen (spatial localization) as shoppers interact with items on shelves. This competition advances computer vision research in retail analytics, with applications in inventory management, customer behavior analysis, and automated checkout systems.
Task Overview Participants will work with short video clips from overhead cameras in retail environments. Each clip is centered around potential shopper-shelf interactions and captured from two synchronized viewpoints (rank0 and rank1 cameras). The model must: - Detect whether any interactions occur in the video. - Temporally localize the start and end frames of each action. - Spatially localize the interaction point between the shopper and the shelf. - Provide predictions for both camera views.
Note: Action classification (take, put, touch) is not part of the evaluation. Focus on detecting and localizing interactions, not classifying them.
Dataset Characteristics Videos may contain one interaction (most common), no interaction (shopper passes by without touching items), or multiple interactions. Each sample includes two synchronized video files. A unique aspect of this challenge is that spatial accuracy is measured in meters, not pixels, accounting for different camera perspectives and varying distances.
Evaluation Submissions are evaluated using mAP Global (mean Average Precision), which combines both temporal and spatial accuracy across multiple threshold values. A prediction matches a ground truth action if the Temporal IoU meets the threshold, the spatial distance (in meters) is within the threshold, and the prediction confidence meets the threshold.
Submission Format
You must submit a CSV file with exactly two columns: id and prediction. The prediction column must contain a JSON string with the structure including score, start/end times, and spatial coordinates for both camera views.