💻 computer science

The 1st AI Children Challenge

The 1st AI Children Challenge introduces a novel dataset of 2D keypoint sequences for children aged 3–16 to advance computer vision applications in pediatric healthcare by addressing the fine-grained analysis of gait behaviors through two clinically validated tracks: Edinburgh Visual Gait Score scoring and the classification of gait patterns in bilateral spastic cerebral palsy.

Original authors: Boyi Li, Yifan Shen, Houze Yang, Xu Cao, Guojun Yun, Li Gao, Turong Chen, Long Xu, Jianguo Cao, Meihuan Huang

Published 2026-08-04

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Boyi Li, Yifan Shen, Houze Yang, Xu Cao, Guojun Yun, Li Gao, Turong Chen, Long Xu, Jianguo Cao, Meihuan Huang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where a doctor can look at a child walking down a hallway and instantly know if their muscles are moving just right, or if something subtle is wrong. This isn't magic; it's the intersection of computer vision (teaching computers to "see" and understand images) and pediatrics (the medical care of children). For a long time, doctors have relied on their own eyes to spot tiny differences in how a child walks, a process called gait analysis. While this works well for experienced doctors, it's hard to teach a computer to do the same thing because children aren't just "small adults." Their bodies are still growing, their movements are wobbly and unpredictable, and they don't walk with the perfect rhythm of a grown-up. This makes it incredibly difficult for standard AI, which is usually trained on steady, adult walking, to spot the tiny clues that signal a health problem.

This is the story of the 1st AI Children Challenge, a high-stakes competition where computer scientists and doctors teamed up to solve this puzzle. The paper describes how they built a massive new "library" of video data featuring thousands of walking children, ranging from toddlers to teenagers. They challenged the world's best AI teams to build models that could act like expert doctors, analyzing these videos to spot specific walking patterns associated with Cerebral Palsy, a condition that affects muscle movement. The goal wasn't just to say "this kid is walking funny," but to score exactly how they are walking and classify the specific type of difficulty they have, all from a simple video. The paper reveals that while this is a tough job, the new dataset and the competition sparked some very clever solutions that are getting us closer to helping doctors diagnose kids faster and more accurately.

The Big Challenge: Teaching AI to Watch Kids Walk

The authors of this paper wanted to bridge a gap between what human doctors can see and what computers can understand. They realized that existing AI tools are like students who have only studied adult textbooks; they fail when they try to analyze a child because kids move differently. To fix this, the team created the Children Gait Pose Sequence (CGPS) dataset. Think of this dataset as a giant, super-detailed training manual for AI. It contains over 1,100 videos of children walking, recorded at 60 frames per second (which is like taking 60 snapshots every second). These videos capture 110 different patients, aged between 2.6 and 16.6 years old, walking in front of cameras set up to see them from the side and the front.

But a video alone isn't enough. The real magic of this dataset is the "annotations"—the labels that tell the AI what to look for. Human experts, including doctors with decades of experience, went through these videos and marked specific points on the children's bodies (like knees, hips, and ankles) to create a skeleton of movement. They then scored 34 different tiny details of the walk, such as "how high the heel lifts" or "how much the knee bends." They also identified four specific ways children with a type of Cerebral Palsy called "bilateral spastic CP" might walk, giving them names like "Jump Gait" (where the child looks like they are hopping) or "Crouch Gait" (where they walk with bent knees).

The Two Tracks of the Competition

The challenge was split into two distinct missions, or "tracks," for the participants to tackle.

Track 1: The Scorekeeper
The first mission was to act like a strict referee. The AI had to look at a child's walk and assign a score to 34 different parts of their movement. This is based on a clinical tool called the Edinburgh Visual Gait Score (EVGS). Imagine a teacher grading a student's handwriting; the AI had to decide if a specific movement was "normal" or "atypical." The paper notes that this is tricky because the differences are often very small. The team had to predict whether a child's heel lift was normal or not, or if their knee was bending too much. The goal was to get the score as close as possible to what a real doctor would write down.

Track 2: The Detective
The second mission was a classification game. Here, the AI had to look at a child with bilateral spastic Cerebral Palsy and figure out which of the four specific walking patterns they belonged to. It's like looking at a group of people and trying to guess which one is wearing a red hat, which one has a blue hat, and so on, but the "hats" are invisible walking styles. The AI had to spot the subtle, pathological differences that separate a "True Equinus" walk from a "Jump Gait." This was particularly hard because the dataset had very few examples of some of these rare patterns, making it easy for the AI to get confused.

How They Measured Success

The organizers didn't just let the teams guess; they set up a rigorous testing ground. They split the data so that the AI could learn on some videos and be tested on others it had never seen before. To make sure the results were fair, they used two main ways to grade the teams:

Accuracy: How often did the AI get the right answer?
Error Measurement: For the scoring track, they measured how far off the AI's score was from the doctor's score.

They also introduced a "post-evaluation" phase. After the competition ended, the top teams had to share their code. The organizers then ran the code themselves on a standard computer to double-check the results. This was to ensure that no team had results that were just a fluke or that their results were just a fluke.

The Results: What the AI Found

The competition attracted 57 teams from around the world. The results showed that while the AI didn't perfectly mimic a human doctor yet, it made significant progress.

For the Scoring Track (Track 1): The winning team, CCE, used a clever mix of pre-trained AI models (which are like general knowledge bases) and specific rules designed by doctors. They managed to predict the scores with high accuracy. Another top team, Btaji Crew, focused on extracting specific "biomechanical" details, like the exact angle of a joint, before making their guess. The paper suggests that the best approach wasn't just using one big AI model, but combining different methods—like using a "hybrid" approach where machine learning works alongside strict medical rules.
For the Classification Track (Track 2): The winning team, seantangth, used a method called XGBoost (a powerful type of decision-making algorithm) but tweaked it to follow specific medical rules about how the left and right legs should relate to each other. They found that simply letting the AI guess wasn't enough; they had to force the AI to respect the symmetry of the human body.

The paper highlights that the top teams all used strategies to handle the fact that there were very few examples of the rare walking patterns. They used techniques like "cross-track signals," where the AI used the results from the scoring track to help it make better guesses in the classification track. It's like using the score of a math test to help you guess the answer to a history question because the two subjects are related.

What This Means for the Future

The paper concludes that this challenge was a major step forward. It proved that with the right data and the right guidance from doctors, AI can start to understand the complex, messy, and wonderful way children move. The authors suggest that future work should focus on making these models even better at handling different types of cameras and different body shapes. They also note that the next big step is to see if these tools can help doctors diagnose children faster in real hospitals.

However, the paper is careful not to claim that the problem is "solved." The results show that while AI is getting good, it still struggles with the most subtle and rare cases. The authors emphasize that the real value of this challenge is not just the winning algorithms, but the creation of a shared dataset and a community of researchers working together to improve child healthcare. The competition showed that when computer vision meets medical expertise, we can start to build tools that help every child walk a little bit easier.

Technical Summary: The 1st AI Children Challenge – Children's Gait Visual Analysis

1. Problem Definition

The paper addresses the critical gap in applying computer vision and AI to pediatric healthcare, specifically focusing on the fine-grained analysis of children's gait behaviors from video data. While quantitative gait analysis is essential for diagnosing movement disorders like cerebral palsy (CP), current clinical practice relies heavily on subjective visual assessments by pediatricians. Existing automated solutions face significant hurdles:

Adult-Centric Bias: Current foundation models assume walking is a mature, stable, and rhythmic process, failing to account for the high entropy and intra-class variance inherent in developing motor systems (ages 2–16).
Clinical Barriers: 3D gait analysis and Inertial Measurement Units (IMUs) are costly and often impractical for uncooperative pediatric patients.
Lack of Data: There was no existing dataset providing ground truth for children's gait sequences with clinical annotations.

To bridge this gap, the authors launched the 1st AI Children Challenge, featuring the Children's Gait Visual Analysis track. The challenge defines two specific problem tracks:

Track 1 (EVGS Scoring): Automating the scoring of the Edinburgh Visual Gait Score (EVGS), a clinical tool comprising 17 observational items per limb (34 total) used to assess gait deviations. The task is formulated as a binary classification (typical vs. atypical) for each item to mitigate class imbalance.
Track 2 (Gait Pattern Classification): Classifying specific gait patterns in Bilateral Spastic Cerebral Palsy into four pathological categories (True Equinus, Jump Gait, Apparent Equinus, Crouch Gait) plus "Within Normal Limits" (WNL).

2. Methodology and Dataset (CGPS)

The core contribution enabling this challenge is the Children Gait Pose Sequence (CGPS) dataset.

Data Collection: The dataset comprises 1,185 videos (339,236 frames) recorded at Shenzhen Children's Hospital using synchronized smartphone and action cameras capturing both coronal (frontal/posterior) and sagittal (lateral) views.
Subjects: 110 pediatric patients (ages 2.6–16.6 years, mean 8.40 years) with a balanced gender distribution (58.2% male).
Annotation Pipeline:
- Pose Estimation: 2D keypoints and bounding boxes were generated using Sapiens-2B and SAM3, respectively.
- Human Verification: All annotations were manually adjusted by human annotators.
- Clinical Ground Truth: EVGS scores and CP motor types/gait patterns were annotated by experienced pediatricians and reviewed by a senior author with 40 years of clinical experience.
Data Split: To prevent identity leakage, the dataset was split at the patient level. The test sets for Track 1 and Track 2 were randomly selected but ensured to be nearly balanced in terms of positive/negative samples and gait pattern distributions.

Evaluation Protocols:

Track 1: Evaluated using Accuracy and Root Mean Square Error (RMSE) on the 34 EVGS items. The final score ( $S_1$ ) combines Accuracy and normalized RMSE.
Track 2: Evaluated using Accuracy and Macro-averaged F1 Score to handle multi-label classification and class imbalance. The final score ( $S_2$ ) aggregates these metrics equally.
Post-Evaluation: To ensure reproducibility, top teams were required to release code. Final rankings were determined by re-evaluating submissions on local hardware (NVIDIA L40S) using three different data partitions.

3. Key Contributions

First Pediatric Gait Dataset: The introduction of the CGPS dataset, the first large-scale collection of 2D keypoint sequences for children's walking motions with clinical ground truth (EVGS and CP subtypes).
Benchmark Establishment: The definition of two clinically relevant tracks (EVGS scoring and CP gait pattern classification) that challenge AI models to handle the high variability of pediatric motor development.
Community Engagement: The organization of the 1st AI Children Challenge, which attracted 57 teams and over 100 participants, fostering a benchmark for pediatric healthcare AI.
Methodological Insights: The paper documents diverse winning strategies, including:
- Hybrid Approaches: Combining deep learning (e.g., MotionBERT, Transformers) with hand-crafted biomechanical features and clinical rules.
- Cross-Track Learning: Utilizing EVGS predictions from Track 1 as priors to inform Track 2 classification.
- Ensemble Methods: Using consensus voting across multiple pose estimation models or classifiers to improve robustness.

4. Results

Participation: 57 teams from around the world participated. 7 teams completed the post-evaluation phase with technical reports.
Top Performers:
- Track 1 (EVGS Scoring): The team CCE achieved the highest average score (0.77817) using a multi-view pipeline combining MotionBERT with PCA compression and clinical descriptors. Btaji Crew and seantangth followed closely, utilizing biomechanical descriptors and hybrid rule-based ML approaches, respectively.
- Track 2 (Gait Classification): The team seantangth won with an average score of 0.52572, employing an XGBoost classifier refined with clinical constraints (Rodda & Graham rules) and symmetry rules. CCE and Zihan Ji followed, leveraging cross-track knowledge and transfer learning from Track 1 backbones.
Performance Gap: The results highlighted the difficulty of the task, particularly in Track 2 where class imbalance and subtle pathological variations led to lower F1 scores compared to Track 1.

5. Significance and Claims

The paper claims that the 1st AI Children Challenge represents a significant milestone in advancing applied computer vision for child healthcare. Its primary significance lies in:

Bridging the Gap: Providing the first available dataset and ground truth to enable AI models to move beyond adult-centric assumptions and address the specific complexities of pediatric gait.
Clinical Relevance: The challenge tasks are directly derived from clinical tools (EVGS) and diagnostic needs (CP subtyping), ensuring that AI advancements have immediate potential for real-world medical diagnosis and rehabilitation.
Future Direction: The authors suggest that future iterations should focus on automated severity assessment, handling transitional gait patterns, and expanding to longitudinal data to evaluate gait evolution. They emphasize that the continued convergence of computer vision, biomechanics, and medical expertise is essential for driving innovation in precision pediatric healthcare.

The paper concludes modestly, noting that while the challenge pushed the boundaries of fine-grained analysis, significant research opportunities remain in handling subtle pathological cues, high inter-patient variability, and generalization to diverse clinical conditions.