Computer vision engineer role: data, failure slices and deployment

Cameras decide more than models do

A model trained on crisp daylight photos will disappoint on a dusty factory camera at night. Much of the job is understanding the capture setup before any training begins: camera position and angle, resolution, frame rate, lighting changes across shifts, lens dirt, motion blur, and compression from the video recorder. Engineers who visit the site, or at least ask for a day of raw footage across all shifts, avoid weeks of rework.

In India, practical variety adds to this: handwritten and printed text in several scripts on the same document, number plates in inconsistent fonts and conditions, and outdoor scenes with harsh sun and monsoon rain. None of these is exotic; they are the normal input.

The artefact: a failure-slice review

Aggregate accuracy hides the failures that matter. After each model version, a CV engineer groups the test set into slices by condition and reviews each. Here is a hypothetical review for a defect detector on a packaging line, where the model must flag torn seals.

Two things make this useful. First, the headline 85% recall would have looked acceptable, but the new-colour slice at 65% would cause missed defects on the line that is about to launch. Second, several fixes are not model changes at all: a lens polariser and a shorter exposure setting belong to the plant's engineering team. The engineer's job is to find these and ask for them clearly.

Recall here is defects found divided by defects present. Compute the combined row from counts, not by averaging the four percentages: 144 + 81 + 47 + 26 = 298 defects found out of 350, or about 85%. False alarms combine the same way, weighted by images. People who make shipping decisions read these reviews, so every combined number must reconcile with its slices.

A realistic sprint

• Two days labelling guideline fixes and spot-checking a vendor's latest batch of annotations.
• Fine-tuning a detector on the enlarged dataset and tracking runs.
• Building the slice review above for the new version.
• Converting the model to a lighter format and measuring speed on the edge device the plant uses, since a model that processes eight frames a second is useless on a line needing twenty.
• A call with the plant team about camera placement for the second line.

Slice recall, frames per second and field stability

• Performance on the slices that matter to the customer, agreed in advance, not only overall accuracy.
• Latency and throughput on target hardware.
• Stability after deployment: does performance hold as lighting, products and cameras change?
• Speed of diagnosis when the field reports misses.

Annotators, embedded engineers and plant teams

Usually an ML or CV lead, within an AI team or a product engineering group. Collaborators: annotation vendors or in-house labellers, backend and embedded engineers, MLOps engineers who run deployment and monitoring (see the MLOps engineer guide), product managers and the customer's operations staff.

Confused with image processing engineer and ML engineer

An image processing engineer often works with classical techniques such as filtering, thresholding and geometry, sometimes close to camera hardware. A CV engineer mostly works with learned models, though good ones use classical tricks when simpler. A general machine learning engineer may serve tabular or text models and handles more of the serving infrastructure.

Jetson, ONNX and OCR: vision posting signals

• "Object detection, segmentation, OCR" — the core task families.
• "Edge deployment, Jetson, quantisation, TensorRT, ONNX" — on-device performance matters.
• "Video analytics, multi-camera, tracking" — video pipelines and identity across frames.
• "Dataset curation, annotation QA" — you will own the data, not just receive it.
• "Research, publications" — a research-leaning team; ask how much reaches production.

Slice-based evaluation before a new architecture

Month one: reproduce the current model's metrics and read every guideline. Month two: add slice-based evaluation for the conditions customers care about. Month three: ship one improvement with a before/after slice table. Next roles: senior CV engineer, CV or perception lead, ML platform roles, or applied research. Slice-based evaluation is not specific to images, and the Data Science programme teaches it with tabular and text data first, which is a sound base before specialising in vision.