Computer vision

Begin by deciding what the system should return. Classification gives an image-level label; detection identifies objects and their locations; segmentation assigns categories or object identities to regions or pixels. These are related tasks, but their outputs are not interchangeable. For example, a label such as “bicycle” does not say where the bicycle is, while a box gives a coarse location and a mask traces its visible region. Choose the output that supports the next decision or action. The COCO dataset illustrates how detection, segmentation, and captioning can be studied as distinct tasks. Its task examples are a useful way to compare what each output represents.

• COCO - Common Objects in Context

  The dataset provides examples of object detection, segmentation, and image captioning tasks.

Computer vision

Computational methods that extract information from images or video.

Object detection

Finding and labeling objects and their locations in visual data.

Segmentation

Assigning visual regions or pixels to meaningful categories or objects.

OCR

Converting text visible in an image into machine-readable text.

A video is a sequence of frames, so a task may ask for an action label across a clip rather than a label for one still image. The Kinetics paper frames human-action classification as a video task and distinguishes it from temporal localization, which asks where in an untrimmed video an action occurs. Document extraction has a different structure: OCR involves finding text and transcribing it, while document understanding can also locate page regions and extract information such as an address. This makes the task definition especially important: transcribing words, finding a table, and extracting a field are separate goals. Use the desired output to decide which stages need evaluation.

• A Survey of Deep Learning Approaches for OCR and Document Understanding

  The survey outlines text detection, transcription, document layout analysis, and information extraction.

• Tesseract User Manual

  The manual describes a text-recognition engine that can extract printed text from images.

A metric is useful only in relation to the task and the errors that matter. For classification, accuracy can be misleading on imbalanced data: a system that always predicts the common class may score well while missing the rare class. Precision measures how many positive predictions are correct; recall measures how many actual positives were found. When comparing visual approaches, choose measures that match the required output and the relative cost of false positives and false negatives. Also inspect examples from the intended setting: images or clips that differ from the evaluation data may expose errors that a single aggregate score hides. Practice by naming the mistake that matters most, then deciding which metric reflects it.

• COCO - Common Objects in Context

  Use its detection and segmentation task examples to think about how the required output shapes evaluation.