Project Overview

The Smart Vertical Basketball Streaming project aims to develop an offline, automated system that converts standard horizontal basketball footage into optimized vertical-format videos suitable for reels, highlights, and mobile viewing. Instead of relying on cloud services or real-time streaming infrastructure, the offline version processes recorded footage locally using computer vision and machine learning techniques.

The system ingests full-game horizontal video and applies player detection, ball detection, and intelligent reframing to identify the most relevant areas of action. Using models such as YOLO for object detection, ByteTrack for persistent multi-player tracking, and Gaussian smoothing with outlier rejection for stable ball trajectory, the system continuously determines the optimal vertical crop that keeps key players and the ball in frame. The pipeline generates a final vertical 9:16 video that preserves the most meaningful gameplay moments without manual editing.

This offline approach ensures faster iteration, reproducibility, and independence from network constraints, making it ideal for performance testing, benchmarking, and automated content creation on local hardware such as GPUs (e.g., NVIDIA A4500) or CPU-based systems. Ultimately, the system enables efficient, high-quality vertical highlight generation directly from traditional horizontal sports footage.

Team Organization

GitHub Repository: Code