Computer Vision Engineer – Sports Video Analytics (Padel)
About the Project
We're building a video analytics platform for padel matches. A single camera
captures the full court, and our ML pipeline processes the footage to detect
players, track them across the match, classify shots, map positions to a 2D
court minimap, and generate match statistics.
The pipeline is working end-to-end. We need someone to help improve accuracy,
fix edge cases, and extend capabilities — not build from scratch.
What You'll Work On
- Improving player identity stability — reducing false identity mints,
cross-side swaps, and gallery contamination after occlusions
- Tuning and extending the Re-ID pipeline (appearance embeddings, cosine
distance thresholds, multi-signal fusion)
- Ball tracking accuracy improvements (small/fast object detection, Kalman
filtering)
- Homography and camera geometry refinement (PnP, temporal smoothing, lens
distortion handling)
- Analyzing diagnostic logs (identity decision logs, cosine distances, track
lifecycles) to diagnose tracking failures
- Potentially training/fine-tuning models on padel-specific data
Required Skills
- Multi-Object Tracking — hands-on experience with BoT-SORT, ByteTrack,
DeepSORT, or similar. You understand track-detection association, track
lifecycle, and occlusion recovery
- Object Detection — YOLO family (v5/v8/v11), Ultralytics ecosystem,
confidence tuning, NMS
- Re-Identification — appearance embeddings, gallery management,
distance-based matching. Experience with OSNet, torchreid, or similar Re-ID
frameworks
- Computer Vision — homography computation, solvePnP, camera calibration,
geometric transforms
- Strong engineering skills — not just scripting. Our pipeline has state
machines, registries, and fusion logic
Nice-to-Have
- Ball tracking experience (TrackNet, heatmap-based detection, Kalman
filtering)
- Pose estimation (MediaPipe, YOLO-Pose)
- Sports video analytics domain experience
- Experience with VLMs/multimodal models for visual understanding
- Audio signal processing (onset detection, FFT)
How to Stand Out in Your Proposal
- Mention specific MOT or Re-ID projects you've worked on
- If you've dealt with identity recovery after occlusion, tell us about it
- Share any experience with sports or broadcast video analysis
- Bonus: explain the difference between BoT-SORT and ByteTrack in one sentence
Contract duration of 1 to 3 months. with 30 hours per week.
Mandatory skills: Computer Vision, OpenCV, Feature Extraction, PyTorch, Convolutional Neural Network