← Back to Dashboard
👁️
Inference Specialization
Computer Vision Engineer
CV engineers build AI systems that interpret images and video — powering autonomous vehicles, medical imaging, facial recognition, manufacturing quality control, and AR/VR. One of the most technically demanding and highest-paying AI specialisations.
40
Days
5
Projects
$185K
Median
↑55%
YoY Hiring
$140K–$250K
Salary Range
Week 1: Advanced Computer Vision Architectures · Days 1–8
Go beyond CNNs: Vision Transformers, modern detection pipelines, segmentation, and 3D geometric understanding.
D1–3
Vision Transformers: ViT, DINO, DINOv2, and SAM
Patch embeddings, attention in image space, self-supervised vision pre-training, Segment Anything Model
↗ EECS 498 Deep Learning for CV — Michigan (YouTube, free)D4–5
Object detection: YOLOv10, DETR, and real-time systems
Anchor-based vs anchor-free detection, DETR attention-based approach, real-time pipelines at 30+ FPS
↗ Ultralytics YOLO documentation (free)D6–7
Semantic and instance segmentation: Mask R-CNN and SAM2
Pixel-level classification, Mask R-CNN architecture, Segment Anything Model 2 API and fine-tuning
↗ Meta SAM2 GitHub and paper (free)
🏗 Project — Day D8: Real-Time Object Detection System
· YOLOv10 fine-tuned on a custom annotated dataset
· Real-time inference at 30 FPS from webcam stream
· mAP at 0.5 IoU evaluation report
· FastAPI deployment with video stream endpoint via WebSocket
Week 2: Contrastive Learning, 3D Vision, and Medical Imaging · Days 9–24
CLIP for zero-shot tasks, 3D point cloud processing, depth estimation, and specialised medical imaging pipelines.
D9–12
Contrastive learning: CLIP, ALIGN, self-supervised vision
Joint image-text embeddings, zero-shot classification, vision-language alignment for downstream tasks
↗ OpenCLIP GitHub and tutorials (free)D13–16
3D vision: point clouds, depth estimation, and NeRF basics
LiDAR data processing, PointNet++, monocular depth with MiDaS, Neural Radiance Fields introduction
↗ Self-Driving Cars Specialization — UoT/Coursera (free audit)D17–20
Medical imaging: DICOM, segmentation, and FDA basics
CT and MRI preprocessing, nnU-Net for automatic segmentation, FDA SaMD regulatory pathway basics
↗ AI in Healthcare — Stanford/Coursera (free audit)D21–23
Video understanding: action recognition and optical flow
3D CNNs, TimeSformer, optical flow with RAFT, action recognition benchmarks on Kinetics-400
↗ PyTorch video tutorials (free)
🏗 Project — Day D24: Medical Image Segmentation System
· U-Net trained on a medical dataset from NIH or Kaggle
· Dice coefficient above 0.85 on held-out test set
· Grad-CAM explanations for 10 clinical predictions
· Model card covering dataset, metrics, and clinical limitations
Week 3: Deployment, Optimisation, and Production Systems · Days 25–40
Deploy vision models at scale: TensorRT acceleration, streaming video pipelines, production monitoring, and autonomous systems.
D25–28
TensorRT and ONNX optimisation for vision models
INT8 calibration for CNNs and ViTs, TensorRT engine building, 3 to 5 times latency reduction
↗ NVIDIA TensorRT documentation (free)D29–32
MLOps for vision: embedding drift and data quality
Pixel-level drift detection, embedding drift with Evidently, monitoring production vision pipelines
↗ Evidently AI for vision monitoring (free docs)D33–36
Autonomous vehicle perception stack and sensor fusion
LiDAR and camera fusion, BEV representations, SLAM for localisation, real-time sensor processing
↗ Self-Driving Cars — University of Toronto/Coursera (free audit)D37–39
Synthetic data for CV: diffusion-based augmentation
SDXL for synthetic training data, domain randomisation, GAN-based augmentation for scarce real data
↗ HuggingFace Diffusion Models Course (free)
🏗 Project — Day D40: Production Vision AI System — Full Capstone
· End-to-end pipeline: ingest, process, detect or segment, serve
· TensorRT-optimised inference under 10ms per frame
· Streaming video WebSocket API endpoint
· Evidently embedding drift monitoring setup
· Architecture write-up and benchmark report