Practical Computer Vision: From Zero to Hero over 10 Topics

Master image classification with hugging face transformers on a plant dataset labeled healthy, powdery, and rusty. Load, split, and transform data from Kaggle, then preprocess with image processors for training.

Model and train an image classification model using a pre-trained auto model, define checkpoint and id2label mappings, and tune training arguments for accuracy.

Push your trained plant-disease model to Hugging Face hub, evaluate with a confusion matrix, and test a Gradio-based interface that classifies uploaded leaf images as healthy, powdery, or rusty.

Convert a trained model to ONNX, explain the interoperable Open Neural Network Exchange format, then run it with ONNX Runtime and deploy via FastAPI.

Create and deploy an image inference api using fast api and Onnx runtime, loading the model at startup, preprocessing images, and serving predictions via post endpoints.

Explore practical computer vision concepts through a tennis video, applying foundational ideas from the course to analyze motion, scenes, and player actions.

Train a YOLOv8 model to detect a tennis ball using a Kaggle dataset, and learn the YOLO label format (class, x center, y center, width, height) with the Ultralytics library.

Install kaggle and ultra linux, download and unzip dataset, and configure yaml. Train the yolo v8 x model with 100 epochs, 800 image size, albumentations, and automatic mixed precision.

Run inference on a single image by loading a test frame, defining a YOLO model with the best weight, and thresholding detections by confidence to locate the ball.

Apply YOLO-based inference to a full video by processing each frame, detecting the ball with bounding boxes, and writing an annotated output video using OpenCV.

Explore zero-shot object detection with grounding dino, enabling open-set recognition via prompt-driven localization, multi-modal reasoning, and contrastive and localization losses to detect novel categories.

Learn how to prompt the DINO model for zero-shot object detection to obtain bounding boxes, using transformers auto processor and auto model, process an image, post-process outputs, and visualize results.

Define tennis player detections on a full video, process each frame, generate predictions with bounding boxes and scores, and prepare for tracking.

Create a global and per-frame tracker to record players, compute Euclidean distances from x mean and y mean, and filter to identify the top two movers.

Project a tennis match onto a reference plane using homography, build the projection matrix from four points with OpenCV, and overlay players, ball, and court lines.

Learn how to compute a homography to project a pitch onto a plane and project the player and ball positions using OpenCV's perspective transform to annotate video frames.

Train a pose-estimation model with ultra and YOLO 11 nano, specifying data, epochs, and image size, then run inference on test images to extract four key points for projection.

Install the diffusers library to access diffusion models and build an image-to-image pipeline with runway ml stable diffusion 1.5. Tune strength and guidance scale to balance prompt adherence and fidelity.