Day 70: Segmentation checkpoint: person cutout pipeline deployed
Second checkpoint: cutout in production
Deploy the segmentation stage: image → person mask → clean cutout, callable as a service and integrated with detection. You now have three pipeline stages live (detect, pose, segment) — each deployed and measured as you built it. This incremental discipline means the full assembly on Day 74 is wiring together tested components, not debugging a big-bang integration.
Deploy and integrate the cutout stage
Ship a /segment (or fold it into a unified endpoint) that returns a person mask/cutout, wired to run after detection. Verify the cutout is clean on varied backgrounds and record its latency. Note which model you're serving in production — SAM for robustness, or your distilled U-Net for speed — and why; that decision is Day-88 interview material.
Track the growing latency budget
Three stages now, each adding latency. Sum them — detection + pose + segmentation — and you have the pipeline's current cost. If it's already heavy on CPU, that's useful early signal for Day 89's performance pass and Stage 6A's optimization work. Measure now; you can't optimize what you never baselined.
Key terms
- Cutout pipeline
- The deployed detection → segmentation path producing a clean person cutout.
- Cumulative latency
- The sum of each stage's latency, giving the pipeline's total time cost.
With three pipeline stages now deployed, why sum their individual latencies?
