HoopCut

HoopCut turns hours of hoops footage into instant highlights. Our AI detects makes, misses, and shot types, auto-clips key moments, and delivers analytics to help players improve and save time.

Comment

🏀 | Inspiration

At HoopCut, we've been living basketball for years- whether that be scoring, watching or playing. We noticed that all players, from semi pro to weekend casuals, spend hours a week watching old footage to edit out clips - wasting hours of valuable playing time. For pros - no clips means no exposure, for casuals - its time lost playing and for leagues - lack of clips means lost marketing opportunities. Phones made capture easy---but editing never caught up. That's the gap HoopCut closes.

🎯 | What it does

  • Instant highlights: Detects made/missed shots and classifies shot type automatically.

  • Auto-clipping: Packages key moments into ready-to-share clips---no manual trimming.

  • Basketball-aware AI: Trained on 5k+ in-game images across courts, lighting, and ball colors; >90% test accuracy (over 1 hour of footage) to detect human, hoop and ball.

  • Player analytics: Shot charts, make/miss splits, tendencies; roadmap for angle-to-FG% insights.

  • Human-in-the-loop: One-tap manual override when needed; your edits improve future results.

  • Capture-agnostic: Film on any device; upload and get highlights. No proprietary hardware.

  • Accessible pricing: Individuals, teams, and leagues---undercutting hardware-locked tools by ~85%.

🧱 | How we built it

  • Vision stack: YOLOv8 + PyTorch + TensorFlow for ball/hoop/player AI model training + trajectory modelling using NumPy to predict makes vs. misses (beyond naïve rim-zone heuristics).

  • Pose signals: MediaPipe Pose Analysis + distance-to-rim heuristics to infer shot types

  • Pipeline: Flask backend for frame-by-frame inference, FFMpeg for editing/exports.

  • AI Stack LabelStudio for data annotation, Google Colab for borrowing GPU Compute

  • Analytics Data analysis done through Numpy, and displayed using MatPlotLib

🧩 | Challenges we ran into

  • Real-world variability: Courts, lighting, occlusion, and ball coloyr broke naive detectors---solved with over 5000 images of custom data across 3 different stadiums, including Pro footage and Casual footage. Changing heuristics based on shot types was a huge challenge. The most challenging of these was three point shots. Determining when a shot was a three, then scaling the hoop size and location to adjust for the physics of longer range shots proved a huge challenge.

  • Video Rewriting Working with different videos, taken on different devices, across different OS was a huge challenge. We solved this by reading metadata and rewriting it for every video.

  • Responsive Design Getting the website to function on MacOS and on large monitors in such a short time span was very challenging.

  • Beyond ~60% heuristics: Simple "activity near rim" methods underperform; motion/angle modelling and per frame ball detection pushed accuracy above 90%.

  • Finding clean data To train the model on footage outside our own games, we had to find clips online which were suitable for annotation. To do this, we had to build multiple scrapers to scrape youtube and NBA highlights.

  • Training AI Not only was data annotation extremely time consuming, we lacked the GPU Compute to train over 5000 images. To solve this, we used google collab and borrowed cloud GPU Compute.

  • Human trust: Users wanted control---shipped manual override within hours of pilot feedback.

  • Privacy & compliance: Built user-controlled deletion and data minimization into the flow.

🎉 | Accomplishments that we're proud of

  • Working prototype: End-to-end upload → detect → auto-clip → export on real games using custom trained AI model.

  • More than a GPT Wrapper Able to undercut OpenAI image detection prices by 3500x using a custom trained AI hosted in the cloud.

  • Accuracy that matters: >90% testing accuracy on over 2 hours of live game footage.

  • On-court validation: Piloted with 40+ players from local leagues to Saturday runs; strong feedback.

  • Fast iteration: Shipped manual labelling/override overnight after first trials.

  • Clear value Similar outcomes to hardware-locked platforms---without $1,600/yr price tags or proprietary cameras.

  • Market Segmentation Analysed the market and identified points of concern, and expressed it through a beautiful pitch video.

🧠 | What we learned

  • Speed + control win: Users want instant results and a way to fix edge cases---manual override became a must-have after pilots.

  • Domain data matters: Every shot taken needs their own specific heuristic, that only basketball players know.

  • Trust is king: Clear deletion/export and minimal data retention build adoption.

  • OpenAI is not the answer Specific problems require specific solutions. OpenAI image detection is not only less accurate, but also 3500x more expensive.

  • Visuals Creating powerpoints that go beyond a white screen with text to get our point across.

  • Teamwork How to effectively communicate ideas across the web.

🌎 | What's next for HoopCut

  • Cloud inference at scale: Autoscaling, priority queues, and speed tiers for fast turnarounds.

  • Mobile-first UX: Upload → review → share flow, lightweight editing, optional captions/subtitles.

  • Deeper analytics: Angle/distance shot charts, release metrics from pose, trend reports; track pilot KPIs (acceptance rate, time saved).

  • Model upgrades: Multi-human tracking, better occlusion handling, temporal action recognition; explore on-device edge inference.

  • Teams & leagues: Coach dashboards, shared libraries, roster profiles, role-based access.

  • Integrations: One-tap publish to YouTube/TikTok, cloud drives, camera vendors; API/webhooks.

  • Go-to-market: University/social league pilots (e.g., Monash, ABA), referral program, watermark-driven virality.

  • Pricing iteration: Post--free-trial tuning, student discounts, partner bundles.

Share this project:

Updates