When a computer vision system disappoints in the field, the cause is often outside the network architecture: a camera remounted at a new angle, three annotators who disagree about what counts as a scratch, or a model that manages two frames per second on the device it must ship on. Computer vision engineers are hired to own that whole chain, from lens and lighting to the final decision. Here is what the work involves, which skills Indian employers want, how interviews test them, and which projects hold up.
What computer vision engineers work on day to day
The core tasks are detection (finding and boxing objects), segmentation (labelling pixels by class or by individual object), tracking (following objects across video frames), and OCR (reading text from images and documents). You work with annotation teams, domain experts such as quality engineers or radiologists, embedded engineers who own the device, and a product team that decides what a wrong answer costs.
A factory inspection project, Monday to Friday:
- Monday: false rejects jump on one line after maintenance replaced a light fitting. You compare brightness histograms from before and after, adjust exposure with the line engineer, and add an automatic lighting check.
- Tuesday: two annotators disagree on whether faint marks are defects. You add example images to the labelling guideline and send the disputed batch back.
- Wednesday: you retrain, then evaluate per defect class and per camera, because one overall number can hide a failing camera.
- Thursday: you export to ONNX and build a TensorRT engine for the edge device. One custom layer is unsupported, so you replace it with standard operations and re-check accuracy.
- Friday: you measure latency from frame capture to the reject signal. The bottleneck is image decoding, not the model.
Next to an ML engineer, you spend more time on data collection, labelling, optics, and hardware limits, and less on tabular features and batch pipelines. Next to an embedded software engineer, you own the model and vision pipeline while they own firmware, drivers, and the operating system; the roles meet at deployment. Vision-language models can now handle some tasks, such as reading a form, without any training, so part of the job is knowing when a small specialised model is cheaper, faster, and more predictable.
The vision toolkit employers ask for
Python, PyTorch with torchvision, and OpenCV are the baseline. Beyond that, the tools map to stages of the pipeline:
| Tool or skill | Where it earns its keep |
|---|---|
| OpenCV | Pre-processing, camera calibration and undistortion, and classical methods such as thresholding and contours that solve some problems with no training at all. It loads images in BGR order, a classic source of silent accuracy loss for models trained on RGB. |
| Detection and segmentation families: YOLO-style detectors, Faster R-CNN, U-Net, Mask R-CNN | Knowing their speed and accuracy trade-offs, not just one repository’s training command. |
| Trackers such as SORT or ByteTrack | Turning per-frame detections into stable identities, so each vehicle is counted once. |
| OCR engines such as Tesseract or PaddleOCR | Text detection and recognition, with document-specific clean-up before and validation after. |
| CVAT or Label Studio | Annotation with review workflows and written guidelines. |
| ONNX, TensorRT, OpenVINO | Portable export, then fast inference on NVIDIA GPUs and Jetson modules or on Intel hardware. |
| GStreamer or DeepStream, plus C++ | Decoding and batching many camera streams; C++ is a common requirement in automotive and edge roles. |
Two topics deserve real understanding. Quantisation: post-training INT8 quantisation uses a small calibration set and is quick, but can cost accuracy, so check hard cases such as small or low-contrast objects; quantisation-aware training simulates low precision during training and usually recovers more. Measure accuracy on the target device, not your laptop. Dataset quality: clear class definitions, rules for occluded and truncated objects, agreement checks between annotators, and splits by video, site, or day rather than by frame, because near-identical neighbouring frames in both training and test sets inflate every metric.
uHired’s market snapshots on role and location pages give the share of current postings that name tools such as OpenCV, TensorRT, ONNX, and YOLO.
Industries hiring computer vision engineers in India
| Industry | Typical problems | What is different |
|---|---|---|
| Automotive and ADAS | Lane, vehicle, and pedestrian detection; driver monitoring | Functional safety processes such as ISO 26262, C++, heavy validation and traceability |
| Retail | Shelf monitoring, footfall counting, checkout assistance | Every store has its own layout and lighting; shopper privacy |
| Manufacturing inspection | Defect detection, measurement, counting | Very few defect examples, tight cycle times, integration with line controllers |
| Medical imaging | Segmentation and triage on X-ray, CT, retinal, or pathology images | Expert labels, clinical validation, and regulatory approval before use |
| Surveillance and smart infrastructure | Number plate recognition, crowd counting, intrusion detection | Legal and ethical limits on identifying and tracking people |
Systems that identify or follow people raise questions to ask before you join the project: what consent or other legal ground covers the footage, how long it is kept, who can access it, and how accuracy varies across groups of people. Images of identifiable people are personal data under data protection laws, including India’s Digital Personal Data Protection Act, 2023. This is general information rather than legal advice; the official text and your employer’s legal team are the authorities. Declining work that conflicts with your values is legitimate, and asking these questions in an interview signals maturity, not awkwardness.
Seniority and titles across vision teams
0–2 years: train and evaluate models under guidance, write data loaders and augmentations, run annotation quality checks, and export models for deployment; edge-focused teams also expect basic C++. 3–5 years: own a use case end to end, from data strategy and model choice to on-device performance and field failures. 6+ years: design systems across cameras and devices, choose hardware with the product team, set validation strategy and safety documentation, and lead a team.
Titles include Computer Vision Engineer, Senior and Lead, Perception Engineer in automotive and robotics, Deep Learning Engineer, and Imaging Scientist at medical companies. Automotive suppliers and engineering services firms run process-heavy projects with formal reviews. Vision startups in retail analytics, inspection, and video analytics give broad ownership across data, model, and deployment. GCCs of global automotive, medical-device, and semiconductor companies offer perception and imaging work, often alongside hardware teams abroad.
Interview rounds and questions to practise
The usual rounds
A coding round, often NumPy or OpenCV operations rather than abstract puzzles; a fundamentals round on convolutions and metrics; a pipeline design round for a camera use case; a project walkthrough; and a manager conversation. Some teams send a take-home such as training a detector on a small dataset; report per-class results and failure examples rather than one score.
Questions that come up
- A 3×3 convolution with stride 2, padding 1, and 64 filters runs on a 224×224×3 input. What is the output shape, and how many parameters does the layer have? (112×112×64, and 1,792 including biases.)
- Why does a convolutional layer need far fewer parameters than a fully connected layer on the same image, and what assumption does that encode?
- Define IoU. Why might COCO-style mAP, averaged over IoU thresholds from 0.5 to 0.95, mislead you when the client only cares whether a defect is present?
- Your detector has strong mAP but misses small defects. How do you investigate?
- Design a pipeline that reads number plates from 40 cameras at a parking facility on one edge GPU box, including night-time images and duplicate reads.
- The model works in the lab and fails on the factory floor. List likely causes in the order you would check them.
- INT8 quantisation dropped accuracy noticeably. What would you try?
- Coding: implement IoU and non-maximum suppression in NumPy.
On the plate reader, panels look for decoding cost on the edge box, frame sampling, a detector followed by a tracker so each vehicle is read once, OCR with confidence voting across frames, validation against the expected plate format, and a stated retention policy for images.
What a vision resume should show
State the task, the data, the deployment target, and a metric the user cares about rather than mAP alone, and show that you handled data yourself. These bullets are invented examples of the shape:
- Example: “Built a surface-defect detector for a stamped-parts line from 6,000 images labelled in CVAT under a written defect guideline; missed defects in the weekly manual audit fell from 14 to 3 while false rejects stayed within the line’s limit.”
- Example: “Converted a PyTorch segmentation model to TensorRT with INT8 calibration on a Jetson module, raising throughput from 9 to 31 frames per second for a 0.6-point drop in mean IoU, so one device could serve each line instead of two.”
- Example: “Halved the character error rate on scanned delivery challans by adding deskewing and adaptive thresholding in OpenCV before recognition, measured on a 1,000-document test set.”
Common mistakes: a tutorial project (“detected cars with YOLO”) presented as original work; one overall metric with no per-class or per-condition breakdown; no mention of hardware or frames per second; and demo videos cut from the best minutes. Show a failure case and explain it.
Routes into computer vision, and projects worth building
Software and ML engineers should learn image fundamentals and OpenCV before deep learning, because pre-processing bugs are a common time sink for newcomers. Embedded engineers already understand the hardware and need model training and evaluation. Electronics and mechanical engineers in manufacturing or automotive can move in through inspection and automation projects, where knowing the line counts as much as knowing the model. Projects worth building:
- Inspection on a small device. Use a public anomaly-detection dataset such as MVTec AD, train a model, deploy it on a Jetson or Raspberry Pi with a camera, and report frames per second alongside per-class recall.
- Tracking with an error analysis. Run a detector and tracker on a public multi-object tracking benchmark, count identity switches, and explain the three most common causes with clips.
- A labelling study. Label 500 images with a written guideline, have a friend label the same set, measure agreement, fix the guideline, and show how the model changes when trained on cleaner labels.
Adjacent paths: the ML engineer guide for broader production ML, the embedded software engineer guide if the device side interests you, and the AI engineer guide for products built on multimodal model APIs.
Vision openings, and what to do before you apply
See current listings in Bengaluru, Hyderabad, Pune, and across India. A plan for the next seven days:
- Read five postings and note the industry, the deployment hardware, and whether C++ is required.
- Implement IoU, NMS, and the convolution output-size formula from memory, with tests.
- Add per-class metrics, hardware, and frames per second to your strongest project’s README; What to Put on GitHub and in a Portfolio covers presentation.
- Sketch the parking-facility pipeline in 40 minutes, then compare it with the points above.
- Schedule the rest of your preparation with A Four-Week Technical Interview Preparation Plan.