Workshop Structure
📅 Day 1: Foundations of Object Detection and Zero-Shot Inference Core Objective: Establish a strong foundation in modern object detection paradigms, understand the YOLO architectural evolution, and execute out-of-the-box inference.
- Evolution of object detection: From fundamental CNNs to the YOLO family
- The end-to-end vs. versatility architectural debate
- YOLOv10 mechanics: Consistent dual assignments and NMS bottleneck elimination
- YOLO11 mechanics: C3k2 blocks and C2PSA spatial attention modules
- Setting up the environment: CUDA, Python, and PyTorch dependencies
- Configuring the Ultralytics Python API pipeline
- Instantiating pre-trained yolov10n.pt and yolo11n.pt architectures
- Running zero-shot inference on standard proxy datasets
🛠️ Hands-on Lab: Configure the development environment from scratch, instantiate pre-trained models, and run basic zero-shot inference on custom images and videos to observe baseline bounding box performance. 🧰 Tools Covered: Python, PyTorch, Ultralytics API, Google Colab
📅 Day 2: Custom Dataset Formatting and Intermediate Fine-Tuning Core Objective: Transition from pre-trained models to domain-specific applications by formatting custom datasets and initiating custom training loops.
- Understanding the YOLO YAML dataset standard and annotation formats
- Data augmentation strategies for robust feature extraction
- Optimizing gradient flow for faster convergence in YOLO11
- Efficiency versus accuracy trade-offs in YOLOv10’s block design
- Formatting specialized datasets (e.g., MVTec AD for industrial defects)
- Initiating parallel distributed training loops for YOLOv10-S and YOLO11-S
- Tracking live metrics (Loss, Precision, Recall, mAP@50)
🛠️ Hands-on Lab: Format a specialized dataset into the YOLO YAML standard, launch custom training loops for both models, and monitor early epoch convergence and live metric degradation. 🧰 Tools Covered: PyTorch, Weights & Biases (W&B), Python, Google Colab
📅 Day 3: Advanced Edge Deployment and Performance Benchmarking Core Objective: Master production-level deployment by exporting fine-tuned weights to optimized inference engines and conducting rigorous absolute latency benchmarks.
- Why standard mAP alone is insufficient for production machine learning
- Understanding computational graph optimizations in inference engines
- Architectural differences between .pt, .onnx, and .engine formats
- Exporting fine-tuned PyTorch weights to ONNX
- Compiling TensorRT engines to simulate real-world edge device constraints
- Processing a continuous 1,000-frame video stream for robust inference testing
- Capturing strict production telemetry (FPS, Peak Memory, Latency Variance)
- Mapping exact speed-accuracy tradeoffs using comparative visualizations
🛠️ Hands-on Lab: Export fine-tuned PyTorch weights to TensorRT, run a continuous video stream benchmark, and generate publication-ready comparative data visualizations (mAP@50-95 vs. FLOPs).
- 🧰 Tools Covered: ONNX, TensorRT, Matplotlib, Seaborn, Google Colab
