Computer vision engineer

What does a computer vision engineer do?

A computer vision engineer builds software that reads images and video.

The models detect objects, read text, track movement or spot defects. You train them, test them and get them running inside a real product.

In Singapore you find this role in logistics and port operations, manufacturing, retail, healthcare, security and government technology teams. It usually sits in an AI or data team, next to machine learning engineers and software engineers.

You work with product managers, data annotators, operations staff and compliance teams. Images of people are often personal data under the Personal Data Protection Act (PDPA), so privacy checks are part of the job.

What the work involves

• Collect and clean image and video data, and check that labels are correct.
• Train and tune models in PyTorch for tasks such as detection, segmentation or classification.
• Run experiments, compare results and find out why a model fails on certain images.
• Package models in Docker and deploy them to servers, cloud services or edge devices.
• Fix bugs and monitor live models for drops in accuracy.
• Explain results and limits to product and operations teams.

Pay in Singapore

Back to top

Skills a computer vision engineer needs

Back to top

Computer vision engineer job description

We are looking for a computer vision engineer to build and deploy models that analyse images and video.

You will work in a small AI team and take models from first experiment to production, then support them after launch. You should be comfortable with Python and PyTorch, and able to explain results to colleagues who are not technical.

Responsibilities

• Build and train computer vision models for detection, classification, segmentation or tracking.
• Prepare datasets, define labelling rules and work with annotators to keep label quality high.
• Design features and data augmentation methods that suit the business problem.
• Run experiments and report accuracy, speed and failure cases clearly.
• Package models in Docker and deploy them to cloud, on-premise or edge hardware.
• Build pipelines that process large volumes of images and video in real time or in batches.
• Monitor live models, find bugs and retrain when performance drops.
• Use pretrained models, for example from Hugging Face, and adapt them when it saves time.
• Follow data protection rules when handling images of people, and record how data is used.
• Document work and present results to product, operations and engineering teams.

Requirements

• Degree in computer science, engineering, mathematics or a related field. Some employers accept strong project work instead.
• Typically 1 to 3 years of experience in machine learning or computer vision for mid-level roles. Entry roles accept graduates with solid projects.
• Strong Python skills and hands-on experience with PyTorch or a similar framework.
• Working knowledge of image processing, convolutional networks and transformer-based vision models.
• Experience with Docker, Git and a cloud platform. Knowledge of model export and optimisation is a plus.

Back to top

Computer vision engineer career path

Most computer vision engineers start as junior engineers on a model or data task.

They move towards owning whole systems and setting technical direction.

1. Junior computer vision engineer

   Entry

   You clean data, train models from existing recipes and fix bugs under guidance from a senior engineer.

2. Computer vision engineer

   Mid

   You own a model from experiment to deployment. You choose methods, run evaluations and handle live issues.

3. Senior computer vision engineer

   Senior

   You design the technical approach for complex projects, review others' work and mentor junior colleagues.

4. Lead or principal computer vision engineer

   Lead

   You guide several projects, set standards for data and evaluation, and work with business leaders on priorities.

5. Head of computer vision or AI

   Head

   You lead the team, manage budgets and hiring, and decide where vision technology fits in the organisation.

The titles are ones commonly used in Singapore; employers may name the levels differently.

Related jobs and their pay

Back to top

Computer vision engineer interview questions

We wrote these questions from the skills this job needs. They are not an official question bank.

1. How does a convolutional neural network learn features from an image, and why does it suit vision tasks?

   What a good answer covers

   • Explains filters sliding over the image and sharing weights
   • Describes early layers finding edges and later layers finding shapes
   • Notes fewer parameters than a fully connected network
   • Mentions pooling or strides to reduce size

2. You are choosing between an object detection model and a segmentation model for a project. How do you decide?

   What a good answer covers

   • Links the choice to the output needed, such as boxes or exact pixels
   • Considers labelling cost, since masks cost more than boxes
   • Considers speed and hardware limits
   • Suggests starting with the simpler option and testing it

3. Which metrics would you use to evaluate an object detector, and what do they miss?

   What a good answer covers

   • Names precision, recall, IoU and mean average precision
   • Explains how the IoU threshold changes results
   • Says averages hide weak classes or conditions such as low light
   • Suggests reviewing failure cases by hand

4. Your model overfits a small dataset. What steps do you take?

   What a good answer covers

   • Checks for leakage between train and test images
   • Adds augmentation such as crops, flips and colour changes
   • Uses transfer learning from a pretrained model
   • Applies regularisation or early stopping
   • Collects more data in the areas where it fails

5. When does feature engineering still matter in an era of deep learning for vision?

   What a good answer covers

   • Mentions preprocessing such as resizing, normalisation and cropping to the region of interest
   • Gives cases with little data or strict hardware limits
   • Notes that hand-made features can feed a lighter model
   • Describes testing whether a feature adds value

6. Describe how you would fine-tune a pretrained model from Hugging Face on your own image dataset.

   What a good answer covers

   • Picks a model suited to the task and licence
   • Freezes early layers first, then unfreezes gradually
   • Sets up a training loop with a proper validation split
   • Tracks experiments and compares against a simple baseline
   • Checks the model licence before commercial use

7. How would you package and deploy a PyTorch vision model with Docker?

   What a good answer covers

   • Pins library versions and the base image
   • Exports the model to a format such as TorchScript or ONNX
   • Keeps the image small and handles GPU drivers
   • Adds a health check and logging
   • Tests the container with sample images before release

8. Your PyTorch training script runs slowly on a GPU. How do you find the cause?

   What a good answer covers

   • Profiles the code before changing anything
   • Checks whether data loading is the bottleneck
   • Tries mixed precision and larger batch sizes
   • Uses more data loader workers and caching

1. A model that counts containers in port yard images works well in daytime but fails at night and in heavy rain. What do you do?

   What a good answer covers

   • Splits results by condition to confirm the gap
   • Collects or simulates more night and rain images
   • Adds augmentation for brightness, noise and blur
   • Considers a different camera setup or infrared
   • Sets a confidence threshold that sends doubtful cases to a person

2. A retailer wants a camera system to count shoppers in its Singapore stores. How do you scope it?

   What a good answer covers

   • Asks what decision the count supports
   • Proposes counting without storing faces or identities
   • Raises privacy and notice requirements with the legal team
   • Defines accuracy targets and a pilot in a single store
   • Plans camera placement and testing in busy periods

3. Accuracy on a live factory defect model drops over three weeks. How do you investigate?

   What a good answer covers

   • Checks for changes in lighting, cameras or products
   • Compares live input images with the training data
   • Reviews logs and a sample of wrong predictions
   • Retrains with new labelled data and tests before release
   • Sets up monitoring to catch drift sooner

1. Tell us about a model that did not work as expected. What did you learn?

   What a good answer covers

   • Describes the problem and the evidence clearly
   • Explains what you tested and in what order
   • Owns any mistake without blaming the data or the team
   • States what you changed in your process afterwards

2. How do you explain a model's limits to an operations manager who wants to launch next week?

   What a good answer covers

   • Uses simple words and examples from real images
   • Gives numbers on error rates and what each error costs
   • Offers options such as a limited launch with human review
   • Agrees on a clear review point

1. You are building a system that reads CCTV footage in a Singapore condominium. What PDPA points do you raise before starting?

   What a good answer covers

   • Asks for a clear purpose and whether notice is in place
   • Limits what is stored, and for how long
   • Considers whether faces need to be identified at all
   • Restricts access to footage and records who sees it
   • Involves the organisation's data protection officer

2. A face-matching model will be used for staff entry in Singapore offices. How do you check that it works fairly across the local population?

   What a good answer covers

   • Notes that the population includes Chinese, Malay, Indian and other groups
   • Builds a test set covering skin tones, ages, glasses and head coverings
   • Reports accuracy for each group as well as overall
   • Sets a fallback such as a card or a security officer
   • Agrees on thresholds with stakeholders before deployment

Common questions about a computer vision engineer job

Sources