Broadcast Sports Video Analyzer
Budget: $2 – $8 USD
I need to set up an end-to-end computer-vision workflow that detects specific objects in my video library. The raw footage is already organized; what it still lacks is accurate frame-level annotation, a robust detection model, and an easy way to run inference on new clips.
Here’s what the project looks like from my side. First, every video frame that contains the target classes must be labeled with tight bounding boxes and class IDs. Once the ground-truth dataset is ready, I want a state-of-the-art detector trained—YOLOv5, YOLOv8, Detectron2, or another modern PyTorch/TensorFlow solution is fine as long as the mAP holds up in validation. After training, the model should be optimized for real-time inference (TTA off, ONNX or TensorRT export where possible) and tested on unseen footage to confirm performance.
Deliverables
• Fully annotated video dataset (COCO-style JSON + reference frames)
• Training notebook or script with reproducible environment file
• Trained weights plus an exported lightweight inference model
• Brief report covering metrics, sample predictions, and improvement hints
• Usage guide showing how to run inference on additional videos
I’ll supply the sport game videos and a label schema once we begin. Let’s keep communication clear so every milestone—annotation, training, optimization—lands on time and matches the acceptance criteria above.
Here’s what the project looks like from my side. First, every video frame that contains the target classes must be labeled with tight bounding boxes and class IDs. Once the ground-truth dataset is ready, I want a state-of-the-art detector trained—YOLOv5, YOLOv8, Detectron2, or another modern PyTorch/TensorFlow solution is fine as long as the mAP holds up in validation. After training, the model should be optimized for real-time inference (TTA off, ONNX or TensorRT export where possible) and tested on unseen footage to confirm performance.
Deliverables
• Fully annotated video dataset (COCO-style JSON + reference frames)
• Training notebook or script with reproducible environment file
• Trained weights plus an exported lightweight inference model
• Brief report covering metrics, sample predictions, and improvement hints
• Usage guide showing how to run inference on additional videos
I’ll supply the sport game videos and a label schema once we begin. Let’s keep communication clear so every milestone—annotation, training, optimization—lands on time and matches the acceptance criteria above.