The Surge of Bouldering, The Problem, and Our Purpose
In recent years, the world of climbing, especially bouldering, has experienced a remarkable surge, with a considerable growth rate of 11.7% according to 99Boulders. As a result of the lack of readily available information for newcomers (currently no other real datasets), there is little known to newbies about how to train, how to approach different boulders, or even how to stay safe.
The Solution: Our Framework
To solve this issue, we've created a dataset that encompasses a wide range of climbing routes, replays, and insights from climbers of varying skill levels. We wanted to go beyond stale online sources and dive deep into the field. For that reason, we went beyond our screens into the real world to collect data! Our goal is to analyze physical trends, shed light on the nuances of different routes, and provide real, but hidden information to climbers seeking to enhance their understanding of the sport. In addition, we extend this data to illustrate how climbing is accessible, climbing improves overall fitness, and that pose estimation can help decrease risk while attempting routes.
The dataset not only addresses the technical, numerical aspects of climbing, such as efficiency and route planning, but also delves into an artistic expression of poses on the wall. By analyzing different body patterns, we can improve safety, athletic development, and overall fun through pattern recognition and other relationships apparent in the data.
Our Story
First, our team headed to Stone Co., College Station's local climbing gym to collect the data. Before each route, we captured frames of only the bare holds with no person in the shot. This allows us/others to use computer vision to analyze where the holds are to create a map of the route. Secondly, we captured each of our attempts (successful and unsuccessful) to scale the wall. Through this method, we were able to not only capture multiple different methods to climb the wall, but also vary the attempts due to our own differing skill levels and biometrics. We conquered as many boulders as fast as we could, noting down the grades, the durations, and the different types of holds.
But this wasnt enough.
We then leveraged the help of the other amazing climbers in attendance at the time to gather more recordings, harder routes, and more and more data in order to build a robust and quantity heavy dataset while maintaining both variety and quality. After 4 hours of moving camera stands over, splitting up to cover more boulders, and getting super sore, we returned home with 40+ recordings each with multiple entries for the tabular end of our dataset.
Our Solution
We curated 2 sets of data within this project. One involves a more normal, tabular set that includes each attempt, the number of moves, elapsed time, grade, and much more. In order to retrieve this data, we sat through each video and noted exactly how many moves each person made with their left foot, their right foot, their hands, exactly when they started, when they finished and a ton of other details. After 3 more hours and 90+ data entries, we were ready to proceed to stage 2.
Stage 2 involved deriving features. From the original raw data, we were able to create models to represent the relationships between all of the data we were able to collect such as the link between age and skill or an obvious weight vs. number of maximum pull ups comparison. We illustrate some of these relationships here:
Climbing is for everyone
When we looked at the data and the processed pose videos, we were amazed to see that climbing seems to be a universal sport. By this, we mean that regardless of age, height, weight, or even number of total pullups, climbing ability was mainly determined by experience. Through an XGBoost model, we achieved a 92.3% accuracy determining whether a climber could successfully climb a route. Through this result, climbers may be able to better evaluate whether to attempt certain climbs that may be risky due to a high chance of falling off. When analyzing the features that guide the model, we see that experience far outweighs any of the other factors. The only way to get better at climbing is to climb more! This is great news for any person worried about their genetics or athletic predisposition getting in the way!
Climbing makes you stronger
We also found, that through only the experience (in years) and someone's highest grade climbed, we're able to accurately predict the number of pullups they are able to perform. As pullups are a basic representation of upper body strength, engaging the back, shoulders, biceps, and lats, we conclude that climbing is a good alternative to standard weightlifting practices for competitive athletes, bodybuilders, or average people. To a 2e-7 pull-up error, we're able to correlate a history of climbing to increases in strength as shown below. Not only can people enjoy climbing boulders, but they will also get stronger while doing it!
Climbing is unique
Climbing is all about body positions and how to optimize movements in order to stay balanced and powerful enough to link sections of a route together. But how can you link these sections if you never see yourself move? This dataset and code allows people to visualize how bodies move when on the wall as well as visualize their own movement. In addition, we classify different body positions using unsupervised clustering and combine it with our tabular dataset to generate a sort of rating for different moves. All in all, not only is the visualization mezmerizing, but the skeletal structure data can be utilized in a variety of novel ways.
Results: Our Impact
We hope that ClimbDB has solved a few problems as well as unearthed some key points about climbing:
1. New Climbers
- New climbers should be able to utilize this easy, organized tabular dataset to generate predictions about success rate, strength, route reccomendations using their own biometric parameters. In addition, they should never be afraid to reach out, as seeing just how friendly all of the climbers at Stone Co. were makes it certain that experienced climbers are ready and able to share much knowledge.
2. Climbing on links
- There exists tons of relationships and links between entities in all aspects of climbing. From strength to pullups to experience and success rate. All of this data can be regressed to determine useful data that can be used in tons of different ways.
3. Climbing is cool!
- Being able to analyze yourself performing movements in hindsight is cruicial for improvement in any physical activity. Climbing is no exception and we hope that the abundance of video footage and pose estimations (which took many gpu hours and some $$) come through as both applicable to classification and recommendation algorithms and cool to look at!
4. Climbing can be safe!
- Utilizing our pose estimation methods, one can classify the risk based on the timestamped pose. While we were unable to fully implement the risk recommendation system, the pose algorithm should serve as a stepping stone towards a safer gym as a whole, as people would be more advised on 'good' moves as opposed to 'risky' ones when necessary.
In addition, we have provided all of our data and code to be as publicly accessible as possible. Feel free to checkout our github or our attached files here on Devpost!
What's next for ClimbDB
ClimbDB would benefit from additional data and a larger diversity of climbers. In addition, more specific data about physical characterstics and metrics such as max bench, max squat, max deadlift etc.., could drastically increase the usability of the data. More work on the pose estimation classification remains to be done in order to reach its potential. One idea we had, but ran out of time was a recommendation system based on the holds within reach and current pose. Depending on the 'risk' of the pose and the nearest and best holds, through induction the model could simulate an entire route and suggest different possible solutions to each boulder.
Thanks
Huge thanks for checking out our project!
Log in or sign up for Devpost to join the conversation.