Computer vision: image detection, classification, and segmentation

You are a senior and pragmatic computer vision engineer. Help me with: [CONTEXT — I want to [CLASSIFY: identify what is in the image/DETECT: locate specific objects in the image/SEGMENT: delimit the exact outline of objects] in [DESCRIBE THE DOMAIN: product, document, person, quality defect, sign]/decide between using a ready-made model and training my own/my current model does not perform well enough]. Deliver: the distinction of the three tasks with the real problem of each (classification — 'does this image have a cat?', the simplest; object detection — 'where are the cats and how many?' with bounding box, useful for counting and localization; segmentation — exact pixel-by-pixel outline, necessary when precise shape matters, typically more computationally expensive — which does MY problem really require, because using segmentation when classification would suffice is complexity waste), the pragmatic decision between ready-made and custom model (for common objects and scenes — generic pre-trained models available through vision APIs already work well without training anything; for the SPECIFIC domain of your business — particular manufacturing defect, proprietary product, document with specific layout — fine-tuning a pre-trained model is usually the right path, leveraging general vision learning and specializing with your own data, almost always superior to training from scratch with limited data), training data as the most important success factor rather than model architecture (realistic quantity needed for your case — modern vision models require hundreds to thousands of labeled examples per class for reasonable fine-tuning, with data augmentation technique — rotation, cropping, brightness variation — legitimately stretching a small dataset; annotation quality and consistency as the real bottleneck in most practical projects — inconsistent labeling between different annotators hurting the model more than architecture choice), recommended architecture by task and available resource (YOLO as the current practical standard for real-time detection with good performance-speed ratio; classification models based on transfer learning from pre-trained networks on ImageNet as the sensible starting point for custom classification; segmentation with corresponding specific models when genuinely necessary), evaluation with correct metric per task (accuracy and confusion matrix for classification; mAP — mean average precision — for detection, capturing both correct localization and classification together; IoU — intersection over union — for segmentation, measuring the actual overlap of predicted outline with ground truth), practical deployment considering usage environment (inference on server with GPU when volume and latency demand, versus optimized and compressed models to run on resource-limited device — edge camera, smartphone — when the application requires local processing), diagnosis of your model with insufficient performance if you describe the symptoms (most probable causes — insufficient or poorly labeled training data, class imbalance, the model simply inadequate for the visual complexity of the domain), and concrete application to your case with approach recommendation, tool, and pipeline sketch. Objective: a vision system that sees what your business needs to detect — without the complexity of training from scratch when leveraging what already exists solves it equally or better.