Deep Learning (DL) · Advanced level

Advanced level · Module 3: Computer Vision Beyond Classification · Lesson 4 of 4

Designing Vision Applications

Turn vision capabilities into useful applications for real users and conditions.

In this lesson you will

  • Match vision tasks to real needs
  • Plan data, evaluation and user workflow for a vision application

Key concepts

task choicefield conditionsuser workflowhuman confirmation

The lesson

Start from the user's task: an extension worker needs a count and a recommendation, not a bounding box. Choose classification, detection or segmentation according to what the task needs.

Plan for field conditions — different phones, light, dust and backgrounds — in data collection and evaluation.

Design the workflow around the model: how users take photos, what they see, how they confirm or correct results and what happens when the model is unsure.

Step by step

  1. Interview or role-play a user.
  2. Choose the vision task.
  3. Sketch the photo-to-decision workflow.
  4. Define what happens with uncertain results.
Case study

An offline pest-counting app for extension workers

A composite, illustrative case of designing a vision model around field realities rather than benchmark scores.

Context

An agricultural programme (a composite example) wanted extension workers to count fall armyworm moths on sticky traps by photographing them with their phones, replacing slow manual counts.

Challenge

Most workers had mid-range Android phones, little mobile data and no connectivity in many villages. An early prototype ran in the cloud and was accurate in the office but useless in the field.

What they did

  1. Re-framed the task from classification to object detection and counting, because workers needed counts per trap.
  2. Collected training photos with workers' own phones in different light, including dusty and partly torn traps.
  3. Chose a small detector, quantised it and ran it on the phone, measuring under one second per photo on the cheapest supported device.
  4. Designed the app to show the count with boxes, let the worker correct obvious mistakes and sync results when a connection appeared.

Results

  • Counting time per trap fell sharply, and corrected counts became new training data.
  • Counts on heavily clustered traps remained less reliable, so the app asked workers to confirm those.

Lessons for your own work

  • Design for the device, connectivity and conditions where the model will be used.
  • A model with a human confirmation step is more useful than a slightly more accurate model nobody can run.

Practice activity

Write a design brief for a vision application of your choice.

Expected result: A brief linking the user's decision to the vision task, data plan and workflow.

Check your understanding

  1. Which vision task suits counting insects?

    Show answer

    Object detection.

  2. What should happen when the model is unsure?

    Show answer

    The user is told and asked to confirm or retake the photo.

Taught live by an instructor — online or in person. Free course lessons are for self-study and do not lead to a certificate — see registration options.