MLB_Hackathon_StatCast++_Challenge4

Inspiration -

Baseball analytics has transformed the game, unlocking new levels of performance and strategy. However, advanced Statcast data has remained exclusive, available only in MLB stadiums equipped with expensive, sensor-based tracking systems. This has left amateur players, coaches, and scouts without access to affordable, high-precision analytics, creating a significant gap between aspiring athletes and elite-level insights.

What it does-

Statcast++ is an AI-powered system that redefines how the game is analyzed and experienced. It instantly extracts essential metrics, such as real-time ball and bat velocities, precise distances, and dynamic trajectories, providing unparalleled insights for fans and analysts. What truly distinguishes Statcast++ is its ability to capture the defining moment when the bat meets the ball, transforming raw data into an awe-inspiring spectacle. Whether you are a passionate fan or a performance-driven coach, Statcast++ offers an immersive experience with real-time analytics and captivating moments.

How we built it-

At its core, Statcast++ leverages the power of Computer Vision and Machine Learning to extract Statcast-level metrics from any video. By utilizing the YOLO model, it accurately detects the bat and ball in each frame, tracking their motion to estimate velocities through frame-by-frame displacement. Google Colab was employed to efficiently train our deep learning models, utilizing cloud GPUs to accelerate development without the need for expensive local hardware. Further enhancing the system's capabilities, Vertex AI enabled scalable deployment and real-time analysis, delivering immediate insights as the action unfolds. Our modeling and analysis are grounded in the MLB dataset, ensuring accuracy on a professional level.

Challenges we ran into

Initially, we faced the challenge of not having a dataset that tracked both the ball and the bat simultaneously. Instead, we had separate datasets for each. To address this, we trained two distinct models based on these datasets, created a dictionary to store the detections for each object individually, and then combined them into a single frame, allowing us to produce the final result.

Next, we lacked side-angle or top-angle views of the pitch, which made it challenging to accurately determine the speed and distances of the bat and ball. To overcome this, we measured the angle the pitch made with the x-axis and computed the distance the baseball traveled accordingly. We applied the same method to calculate the bat's movement.

Lastly, we did not have access to local GPUs for training our models. However, we leveraged Google Colab’s on-demand GPUs, which enabled us to achieve state-of-the-art detections in a significantly reduced timeframe.

Accomplishments that we're proud of

We successfully combined multiple training models into a single output when confronted with an insufficient and skewed dataset, ensuring that our analysis remained robust and accurate. Despite not having access to high-end sensors or trackers, we were able to reliably determine the trajectories, velocities, and distances of both the bat and baseball, leveraging innovative solutions to fill the gaps.

Whenever we encountered roadblocks or faced challenges that made certain approaches unfeasible, we pivoted and embraced new ideas, ensuring that the project continued to progress and evolve. This adaptability played a key role in overcoming obstacles and achieving our goals.

What we learned

Data Fusion: By combining separate models for the bat and ball, we learned to integrate distinct datasets into a cohesive output, improving accuracy.

Leveraging Resources: We discovered the value of using tools like Google Colab’s on-demand GPUs, enabling us to scale our models without needing expensive hardware.

Real-Time Analytics: The project demonstrated that accurate performance tracking is achievable without high-end sensors, making advanced analytics more accessible.

What's next for MLB_Hackathon_StatCast++_Challenge4

Incorporating Multiple Camera Angles: By adding additional camera angles, we can better determine the trajectories of the ball and bat, offering a more accurate and detailed analysis.

Engaging Visual Enhancements: Introducing animations and pop-up statistics during key moments can make the video more interactive and engaging for fans, enhancing the viewing experience.

Sensor Integration: Augmenting the model with data from sensors that track ball spin, pitch angle, and bat acceleration will provide even more granular insights, further improving the accuracy of our performance analysis.