| name | computer-vision-pipeline |
| description | Run the computer-vision MLOps lane end to end: data/annotation -> CV task -> architecture choice -> training (transfer-learn first) -> eval by task (mAP/IoU/CER/OKS) -> serving/edge placement. CV-specific leakage (scene-aware split, augment-after-split). Seams back to training/serving/monitoring agents. |
Computer-Vision Pipeline
CV is a sub-domain of MLOps, not a separate track — it runs the same train->register->serve->monitor loop with image/video-specific choices. Full decision trees + notes: ../../knowledge/computer-vision-engineering.md.
1. Data & annotation
Define a label guideline, measure inter-annotator agreement, version the label schema (a relabel = a new data version). Class imbalance is the norm — track per-class, not aggregate. Watch metadata/near-duplicate leakage (capture device, filename, timestamp). Faces/plates/medical/PII -> data-governance-privacy + security-engineering.
2. Task -> architecture
Pick the cheapest family that answers the question (task -> architecture tree): classification, detection (real-time YOLO/RT-DETR vs accuracy-first DETR/Faster-RCNN), semantic vs instance segmentation, OCR, keypoint/pose. Then take that family through the build-vs-fine-tune-vs-prompt sourcing tree.
3. Training
Transfer-learn / fine-tune a pretrained backbone first; from-scratch only for exotic-domain + large-data cases. Freeze-then-unfreeze; version the pretrained checkpoint exactly. CV leakage discipline: split by scene/source/patient, not random frame (near-duplicate frames leak); augment the training split only, after the split, and only with label-preserving transforms (a flip can be label-destroying for text/pose).
4. Eval by task
The metric is task-specific: classification -> precision/recall/F1/PR-AUC (accuracy lies under imbalance); detection -> mAP@IoU; semantic seg -> mean IoU/Dice; instance/panoptic -> mask mAP/PQ; OCR -> CER/WER; keypoint -> PCK/OKS-mAP. Eval set held out and not augmented; per-class/per-slice over aggregate. Whether a lift is real -> applied-statistics.
5. Serving / edge placement
Choose edge-vs-cloud (inference-placement tree): cloud batch (offline corpus), on-device/edge (real-time / no connectivity / raw imagery can't leave), edge-opt (quantize/prune/distill + compile — re-measure accuracy after), edge->cloud hybrid, or cloud real-time. Same preprocessing in train and at the inference target, or you have image-form training-serving skew.
Seams back to the core agents
- Training (reproducible pipeline, scene-aware split, augmentation discipline, experiment tracking) ->
training-pipeline-engineer.
- Serving / edge (placement, quantize/compile + accuracy re-check, latency budget) ->
model-serving-engineer.
- Monitoring (CV input drift = lighting/camera/domain shift; decay; retrain trigger) ->
ml-monitoring-engineer.
- Architecture / build-vs-buy ->
ml-platform-architect.
- LLM/vision-generative or open-ended NL-over-image tasks ->
claude-app-engineering (this team owns classical/custom-CV MLOps).