Computer Vision Engineer

What a Computer Vision Engineer does daily

• Image classification and object detection model development — training and fine-tuning deep learning models; image classification with ResNet, EfficientNet, ViT (Vision Transformer); object detection with YOLO (v8/v9/v10, the most widely deployed detection architecture), DETR, RT-DETR; instance and semantic segmentation with Mask R-CNN, SAM (Segment Anything Model); selecting the right architecture for the task's accuracy/speed/size trade-off
• Dataset preparation and annotation — computer vision is heavily data-dependent; sourcing, cleaning, and annotating image datasets using tools like Label Studio, CVAT, Roboflow, or V7 Darwin; defining annotation schemas (bounding boxes, polygons, keypoints, semantic masks); calculating inter-annotator agreement; data augmentation pipelines (Albumentations, torchvision.transforms) to expand dataset diversity without additional annotation cost
• Transfer learning and fine-tuning — adapting pre-trained CV models to specific domains and tasks; fine-tuning ImageNet-pretrained CNNs on domain-specific data (medical images, satellite imagery, manufacturing defects); understanding when to freeze layers vs train end-to-end; few-shot learning for scenarios with very limited labelled data; the standard production approach since training from scratch is rarely justified
• Classical image processing — OpenCV for pre-processing, filtering, and feature extraction; Gaussian blur, morphological operations (erosion, dilation, opening, closing); edge detection (Canny, Sobel); contour detection; perspective transformation and image warping; colour space conversion (RGB, HSV, LAB, grayscale); classical methods remain important for preprocessing pipelines and for tasks where deep learning is over-engineered
• Model deployment and inference optimisation — converting trained PyTorch/TensorFlow models to deployment formats (ONNX, TensorRT for NVIDIA GPU acceleration, TFLite for mobile/edge, CoreML for Apple devices, OpenVINO for Intel CPUs); inference server deployment (Triton Inference Server, TorchServe, FastAPI-based REST API); optimisation techniques (quantisation, pruning, knowledge distillation) to reduce model size and increase inference speed for edge and real-time applications
• Video analysis and multi-object tracking — processing video streams frame-by-frame; temporal consistency challenges; multi-object tracking (DeepSORT, ByteTrack, BoT-SORT) for tracking multiple objects across frames; optical flow (RAFT, Farneback) for motion estimation; action recognition; anomaly detection in surveillance video; real-time processing pipeline design with OpenCV and GStreamer
• Medical imaging (for healthcare CV) — DICOM format processing (pydicom); working with CT, MRI, X-ray, and fundus photography modalities; adapting deep learning models to 3D volumetric data (3D U-Net); class imbalance handling for rare pathology detection; FDA/regulatory compliance requirements for medical AI; privacy and de-identification of medical images (important for Sri Lankan hospital AI projects)
• 3D computer vision — stereo vision (depth from two cameras); LiDAR point cloud processing (Open3D, PCL); depth estimation from monocular images (MiDaS, DepthAnything); 3D object detection; NeRF (Neural Radiance Fields) for 3D scene reconstruction; photogrammetry and structure-from-motion; relevant for robotics, autonomous vehicles, and AR/VR depth sensing
• MLOps for computer vision — managing CV model lifecycle in production; data versioning (DVC); experiment tracking (MLflow, Weights & Biases); model registry; continuous model evaluation on production data; monitoring for data drift (distribution shift in incoming images); retraining pipelines triggered by performance degradation; the production infrastructure that separates research-grade from production-grade CV systems
• Industrial and edge CV deployment — deploying CV models on edge hardware (NVIDIA Jetson Nano/Orin, Coral Dev Board with Google Edge TPU, Intel Neural Compute Stick); camera hardware selection (industrial GigE cameras vs consumer USB webcams; rolling vs global shutter trade-offs; near-infrared for low-light inspection); integration with PLCs and factory control systems for real-time quality control