The CUHK-X Multimodal Human Activity Challenge is an international competition organized by the AIoT Lab at The Chinese University of Hong Kong. It invites teams from all over the world to advance privacy-preserving human activity recognition, understanding, and reasoning using non-RGB sensing modalities — depth, IMU, millimeter-wave radar, skeleton, thermal, and infrared — built on the CUHK-X benchmark. Teams first compete online on Kaggle across two parallel tracks, and the top teams then advance to the on-site finals at UbiComp 2026 in Shanghai to compete for a total prize pool of USD $20,000.

A Two-Track Challenge

The CUHK-X Challenge is organized as two parallel tracks, each with its own Kaggle competition, leaderboard, and USD $10,000 prize pool:

• Small Model Track — Human Activity Recognition (HAR): build lightweight multimodal models that classify 40 daily activities from depth, IMU, mmWave radar, and skeleton data.
• Large Model Track — VQA for Action Understanding and Reasoning: apply multimodal large models to visual question answering on privacy-preserving videos, covering human action understanding (HAU) and human action reasoning (HARn).

How to participate

1. Register your team on the official challenge website: openaiotlab.github.io/CUHK-X-Challenge. Registration on the official website is required to be eligible for prizes, announcements, and finals invitations.
2. Join your track(s) on Kaggle (Stage 1) using exactly the same team name as registered — Small Model Track (HAR) · Large Model Track (VQA). Teams may enter one or both tracks; certificates and shortlist notifications are matched by team name.
3. Download the data, train your model, and submit predictions on Kaggle before the leaderboard freeze on September 15, 2026.
4. Advance to the selection stage (Stage 2): the Top 15 teams per track on the Kaggle private leaderboard enter an anti-cheating verification stage (September 16–30, 2026), submitting their code and running live inference in a recorded Zoom session; the Top 6 teams per track advance to the finals.
5. Compete in the on-site finals (Stage 3): finalists run inference on a brand-new private dataset at UbiComp 2026 in Shanghai on October 11, 2026, give a 15-minute technical presentation, and attend the awards ceremony.

Awards

Each track carries an independent prize pool of USD $10,000 (USD $20,000 in total). The following prizes apply independently to both the Small Model Track and the Large Model Track:

• 1st Place: USD $6,000
• 2nd Place: USD $3,000
• 3rd Place: USD $1,000

Three special awards will additionally be presented per track (prize amounts to be announced):

• Best Report Award: selected by the review committee for the best technical report.
• Most Popular Award: determined by community vote for the most innovative solution.
• Best Faculty Advisor Award: for the advisor of the highest-scoring student team.

Beyond the cash prizes, every participating team is recognized through a five-tier certificate system (per track). Tiers are nested — each team receives only its highest-qualifying award:

All certificates are issued electronically to the email address provided during team registration.

Travel grants: every finalist team (Top 6 per track) attending UbiComp 2026 in person receives a travel grant of up to USD $500, reimbursed against actual expenses. Teams unable to travel may join the finals remotely via Zoom and remain eligible for all prizes and awards; travel grants apply only to in-person attendance.

Summary of Key Dates

• June 20, 2026: Competition launch — both Kaggle competitions open for registration and submissions; the challenge dataset is publicly released. (already started)
• September 15, 2026: Kaggle leaderboard freeze. The Top 15 teams per track on the private leaderboard are notified and must upload their code and model checkpoint within 48 hours.
• September 16–30, 2026: Selection stage — Zoom-based verification sessions with live inference on freshly released sample data.
• October 1, 2026: Final Top 6 teams per track announced; final technical report due.
• October 11, 2026: Finals at UbiComp 2026 in Shanghai — on-site inference, technical presentations, and awards ceremony.

Competition Format

The challenge consists of two independent parallel tracks, each with its own Kaggle competition, leaderboard, prize pool, and evaluation criteria. Teams may participate in one or both tracks. Both tracks follow the same three-stage structure.

Small Model Track — Lightweight Human Activity Recognition

This track targets resource-constrained edge deployment in smart-home and healthcare scenarios — applications such as Alzheimer’s monitoring, fall detection, and elderly care, where models must run on low-power devices with limited memory and compute. Participants build lightweight multimodal models that fuse depth imagery, IMU streams, mmWave radar, and skeleton keypoints to classify 40 daily activities under a strict cross-subject evaluation protocol.

• Task: 40-class human activity recognition (cross-subject)
• Cross-subject split: training on users 1–9 and 16–24; testing on users 10–11 and 25–26
• Modalities: depth, IMU, mmWave radar, skeleton, infrared, thermal
• Model constraints: conventional architectures (CNN / RNN / Transformer); model size ≤ 100 MB; no large pretrained backbones
• Prize pool: USD $10,000

Large Model Track — Multimodal VQA (HAU & HARn)

This track pushes the frontier of large vision–language models on non-RGB modalities. Participants tackle human action understanding (HAU) and human action reasoning (HARn) through visual question answering on privacy-preserving videos. There is no parameter limit, encouraging exploration of prompt design, modality alignment, and fine-tuning at scale.

• Task: visual question answering on privacy-preserving videos (HAU and HARn)
• Cross-subject split: training on users 1–9 and 16–24; testing on users 10–11 and 25–26
• Modalities: depth, thermal, infrared, skeleton, IMU, and mmWave radar
• Model constraints: no parameter limit; large vision–language models encouraged
• Prize pool: USD $10,000

Final Scoring (Finalists)

The final ranking of the Top 6 teams in each track is determined by:

• Kaggle private leaderboard — 20%
• Final on-site private test — 30%
• Reproducibility (selection stage) — 10%
• Technical report — 20%
• Presentation — 10%
• Model efficiency — 10%

The CUHK-X Dataset

Most large vision–language models still depend almost entirely on RGB data, while modalities such as depth, thermal imaging, IMU, and millimeter-wave radar remain severely underrepresented — largely due to the lack of large-scale, high-quality paired multimodal datasets.

CUHK-X, built by the AIoT Lab at CUHK, addresses this gap with 64,267 fully synchronized samples across seven modalities (RGB, depth, thermal, infrared, skeleton, IMU ×5, and mmWave radar), collected from 30 participants performing 40 daily activities in two real-world indoor environments. Annotations follow a Ground-Truth-First strategy that combines LLM-generated scene descriptions with human review to ensure temporal and logical consistency. The dataset supports three progressive tasks: HAR (action classification), HAU (action understanding), and HARn (action reasoning).

RGB data are excluded from the challenge; all remaining modalities are permitted. The cross-subject split assigns 18 participants to training, 4 to the public test set, and 8 to a held-out private test set.

The track-specific challenge data are available from the following mirrors:

• Small Model Track: Hugging Face · Google Drive · Baidu Netdisk
• Large Model Track: Hugging Face · Google Drive · Baidu Netdisk

Further details are available on the CUHK-X dataset homepage and in the GitHub repository.

Eligibility and Team Rules

Eligibility

• Open to students, researchers, and industry teams worldwide
• Cross-institution and cross-country teams are permitted
• Members of the CUHK AIoT Lab and their direct collaborators are ineligible for prizes
• Participants must comply with Kaggle’s terms of service

Team Rules

• Team size: 1–3 members (a faculty advisor is not counted)
• Each individual may join only one team per track; participation in both tracks is allowed
• Team mergers on Kaggle lock 7 days before the submission deadline

Verification and Anti-Cheating

All Top 15 teams per track must pass a verification session before advancing to the finals. By September 22, 2026 (23:59 UTC), shortlisted teams upload a submission package containing:

• Full training and inference code
• Final model weights (checkpoint)
• A single-entry inference script (data directory in, prediction CSV out)
• A README documenting the reproduction steps
• A signed honor declaration

Each team then attends a recorded 45-minute Zoom session. A link to sample data — containing both seen and unseen subjects — is released at the start of the session, and teams have up to 2 hours to complete inference and submit their results. The committee additionally reproduces each team’s Kaggle result offline from the submitted code. A team passes if the reproduced accuracy is within 10% of its private-leaderboard score; teams that fail are replaced by the next-ranked team.

Conference Presentation

The Top 6 teams per track are invited to the finals at UbiComp 2026 in Shanghai on October 11, 2026. Finalists run inference on a brand-new private dataset on site, and each team gives a mandatory 15-minute technical presentation followed by Q&A. The awards ceremony takes place on the same day.

In-person attendance is strongly encouraged and supported with travel grants. Teams unable to travel may participate remotely via Zoom — organizers will operate the projector and coordinate live Q&A on their behalf — and remain eligible for all prizes and awards.

Competition Rules

• Fair play (both tracks): manual labeling of test samples, use of test-set ground-truth labels in training (in any form), and multi-account registration or collusion between teams are strictly forbidden.
• Small Model Track: no large pretrained backbones, and no closed-source APIs or LLMs — whether for development or for labeling training data.
• Large Model Track: any pretrained model (including large vision–language models) and closed-source APIs are allowed; LLM-based pseudo-labeling and prompt engineering are encouraged.
• IP and code usage (Kaggle standard): participants retain full copyright on their code and models; all competition data remain the exclusive property of the CUHK AIoT Lab under the dataset license. Finalist teams (Top 6 per track) must open-source their solutions under Apache 2.0 within 30 days of the finals or decline their finalist status; non-finalist code is deleted after the competition.