Challenge URL:https://www.prcv.cn/web/#/competitionsmeeting_competition_vision

Open the container, take out a drink, grab two bags of snacks, close the door and leave.

For humans, this is an ordinary purchase. For machines, accurately answering"how many items were taken and what they were" still faces many challenges.

Hand and body occlusion of products, difficulty distinguishing similar packaging, reflections from cabinet lights, fast and continuous grabbing actions, multiple items moving simultaneously... and users may pick up items to examine and then put them back.

How can machines truly understand a complete transaction video?

the main theme of digital economy development during the "14th Five-Year Plan" period, and new-type entity enterprises are typical representatives of this integration, continuously unleashing the value potential of data elements under deep digital-real integration. In the past three years, the connotation and extension of "digital-real integration" have been continuously enriched and rapidly expanded, with the "digital" in "digital-real integration" expanding to include "digital technology" and "data elements."PRCV One of the 2026 competition projects, hosted by CloudWalk Technology, which provides the tasks and dataSmart Vending Machine Transaction Video Recognition Challenge, now openly recruiting participants from universities, research institutes, corporate teams, and individual developers.

10,000 training videos, 1,000 validation videos, and 29,000 RMB in prizes are ready.

Welcome to take on the challenge of video understanding in real retail scenarios.

One Video, Two Questions

Participating teams need to analyze a complete smart vending machine transaction video and output two results:

How many items were taken in this transaction?

What are the specific SKUs of the items taken?What are the SKUs?

The competition only provides raw transaction videos and transaction-level labels, without frame-level detection boxes, item trajectories, action timestamps, or temporal annotations.

Participants can adopt technical approaches such as video understanding, multimodal large models, object detection and tracking, temporal action analysis, rule-based post-processing, or combine multiple methods to complete the challenge.

Why Is It Worth the Challenge?

Smart vending machines are entering scenarios such as unmanned retail, campus supply, and airport/train station convenience services.

Recognition tasks in real environments are far more complex than standard image classification:

· Frequent occlusion by hands, body, and cabinet doors;

· Highly similar packaging for products of the same brand and series;

· Reflections from plastic, glass, and metal packaging;

· Brief effective recognition frames due to fast grabbing;

· Multiple items may move simultaneously;

· Actions like taking, viewing, and returning are easily confused.

The model must not only recognize product appearance but also understand the complete transaction process.

This challenge aims to promote the application of technologies such as video understanding, multimodal large models, fine-grained product recognition, object tracking, and temporal behavior modeling in smart retail scenarios through unified data, tasks, and evaluation, resulting in reproducible, evaluable, and iterable solutions.

Competition Data

This competition provides three types of data:

Training set:10,000 real smart vending machine transaction videos
with open labels for item quantity andSKU, used for model training and solution debugging.

Validation set:1,000 transaction videos
with open labels for item quantity andwith SKU labels, used for local validation, format verification, and self-testing before submission.

Test set: hidden label evaluation
The test set only provides video filenames and download links, without disclosing item quantity orSKU labels to participants; offline evaluation is conducted uniformly by the organizing committee.

The organizing committee will also provide an officialSKU dictionary, label specifications, and submission examples to unify standard product names and result formats.

Award Settings

The total prize pool for this challenge is29,000 RMB.

First Prize1 team: 10,000 RMB
Second Prize2 teams: 5,000 RMB each
Third Prize3 teams: 3,000 RMB each
Several Excellence Awards

Winning teams will receive certificates issued by the competition organizers. Outstanding teams will be invited toPRCV present technical reports, exchange results, and attend the award ceremony during the 2026 conference.

Competition Rules

I. Eligibility and Team Requirements

The competition is open to higher education institutions, research institutes, corporate teams, and individual developers. No registration fee is charged.

Participants may register as individuals or teams. Each team is recommended to have no more than5 members, and must designate one team leader as the primary contact.

Each participant may only join1 team. Repeated teaming, fraudulent submissions, malicious leaderboard manipulation, or any other behavior that affects the fairness of the competition will result in disqualification or voiding of results.

II. Registration Method

Registration email:

prcv2026-smartcabinet@cloudwalk.com

Email subject format:

Smart Vending Machine Transaction Video Recognition Challenge_Team Name

Email body must include:

· Team name;

· Participating organization;

· Team leader name;

· Number of team members;

· Contact phone number;

· Contact email.

It is recommended to name the registration form attachment as:

Smart Vending Machine Transaction Video Recognition Challenge_TeamName.xlsx

Registration deadline:

2026-06-30 23:59:59

Registration becomes effective after the organizing committee confirms the team information and signs the data usage agreement.

3. Competition Format and Data Distribution

This competition adopts a single-round format.

Email will serve as the unified communication channel for registration confirmation, data list distribution, test set notification, result submission, and score review.

Training set, validation set, and test set videos will not be sent as email attachments. The organizing committee will send the team leader via the registration emailan Excel data list and download instructions. Participating teams should download the original videos using thevideo_url in the list.

Download links are for internal use by the participating team only and must not be forwarded, made public, or used for purposes other than this competition.

The test set will be released24 hours before the result submission deadline and will not provide any labels.

4. Task Setup and Technical Rules

The input and output requirements for each test video are as follows:

Input: A single complete transaction video.
Output: Total number of items countand itemSKU listsku.

Participating solutions do not need to output detection boxes, item trajectories, action timestamps, or frame-by-frame recognition results.

The competition allows the use of publicly available pre-trained models and public datasets. Participating teams should honestly disclose the sources of external models, data, and tools in the final technical description.

The competition prohibits the use of:

· Non-public external smart cabinet transaction data;

· Leaked test set data or answers;

· Additional labels that may undermine the fairness of the competition;

· Solutions that manually hardcode answers for the test set.

5. Key Milestones

June 15
Release competition information and open registration

June 30
Registration deadline, qualification confirmation, and data usage agreement signing

July 1
Release training set, validation set, label specifications,SKU dictionary, and submission example

August 5
Release test set video list without providing test labels

August 6
Participating teams submit test set prediction results

August 9
Complete format validation, offline scoring, and abnormal result review; announce competition rankings

August 20-22 PRCV 2026 Conference
Plan to invite winning teams for technical exchange, result presentation, and award ceremony

Each team may submit up to3 test results; the last valid submission before the deadline shall prevail.

VI. Method of Submitting Competition Results

Participating teams must submit a compressed package with the file naming format:

Smart Vending Machine Transaction Video Recognition Challenge_TeamName.zip

The compressed package must contain at least submission.csv, and may optionally include readme.txt to describe the model version and submission notes.

submission.csv must include the following three columns (column names are case-sensitive):

· filename: test video file name;

· count: total number of items taken in this transaction;

· sku: SKU result of items takenSKU result.

Result format requirements:

1. filename must cover all test videos and must not include file names outside the test set;

2. count must be a non-negative integer;

3. multipleSKU values are concatenated with half-width plus sign "+";

4. When multiple items of the sameSKU are taken, they must be repeated;

5. When no items are taken, countshould be0, sku should be NONE;

6. SKU names must use the standard names from the official SKU dictionary;

7. count must be consistent with the number of items after expanding sku.

Submission email:

prcv2026-smartcabinet@cloudwalk.com

Top-ranked teams must, as required by the organizing committee, additionally submit complete engineering code, model weights, inference scripts, runtime environment description, and usage notes for external data and models.

If the submitted results cannot be reproduced, the organizing committee has the right to cancel the team's score and adjust the ranking accordingly.

VII. Scoring Rules

The competition scores from two dimensions: item quantity recognition andSKU recognition.

1. Item quantity accuracy, accounting for60%

Determine whether the predicted count matches the actual number of items for each case, then calculate the average accuracy across all test videos.

2. SKURecognitionF1, accounting for40%

After expanding SKU results by item count, calculate the F1 score using a multiset approach.

For example, the sameIf 2 SKUs are actually taken but the model only identifies 1, a missed detection result will be generated. If both the actual result and the predicted result areNONE, this videoSKU F1 is recorded as 1.

3. Final Score

Final Score = Product Quantity Accuracy × 60% + SKU Recognition F1 × 40%

If total scores are tied, compare in order:

1. Transaction Complete Accuracy, i.e., the proportion of videos where both product quantity andSKU are completely correct;

2. Valid result submission time, with earlier submissions ranking higher.

VIII. Data Usage and Intellectual Property

Competition data may only be used for this competition and related academic research. Redistribution, sale, public upload, or use for commercial product development is prohibited.

Algorithms, code, model weights, technical solutions, and related intellectual property submitted by participating teams belong to the teams. The organizing committee may use relevant materials only for score verification, result reproduction, technical exchange, and competition promotion.

Participating teams must ensure that their submitted solutions and the data, models, and tools used do not infringe upon third-party intellectual property rights. Disputes arising from improper use shall be the responsibility of the participating team.

The organizing committee has the right to review abnormal submissions, suspected data leaks, manual hardcoding of test set answers, etc., and disqualify violators. If a team disputes the results, they may submit a review request within the time specified by the organizing committee.

IX. Contact Information

Contact Person:Liu Shengzhong

Contact Phone:18580088822

Contact Email:liushengzhong@cloudwalk.com

Competition Registration and Result Submission Email:

prcv2026-smartcabinet@cloudwalk.com

Let machines truly understand a shopping trip

Every opening, taking, returning, and closing of a smart cabinet door tests visual perception, product recognition, temporal understanding, and reasoning capabilities.

10,000 real transaction videos are ready.

What did he take?

The answer awaits you.