Day 88: Defending model choices: detection vs segmentation vs parsing
Interview rehearsal: defend every choice
A Stage 2 exit criterion is being able to explain detection vs segmentation vs parsing and defend each model choice. This is interview rehearsal. For each component, you should crisply state: what it does, why you chose that specific model, what you rejected and why, and its trade-offs. Being able to *defend* choices — not just list technologies — is what separates a senior candidate from a tutorial-follower.
- Detection (YOLO): "One-stage for speed on a real-time-ish path; Ultralytics tooling; accurate enough for person+garment. Rejected Faster R-CNN — accuracy edge not worth the latency here."
- Pose (MediaPipe): "CPU-real-time keeps the pipeline on the ₹0 droplet; 33 landmarks with visibility for honest confidence. Fine for single front-on subjects."
- Segmentation (SAM / distilled U-Net): "SAM for robustness over GrabCut's color-based failures; a distilled U-Net when speed matters. I built U-Net to understand it — and it's Stage 4's diffusion backbone."
- Parsing (SCHP): "Region-level detail for anatomically-targeted measurement and multi-source fusion, covering pose's blind spots."
The strongest answer names the trade-off
Anyone can say 'I used YOLO'. The senior answer is 'I used one-stage YOLO because latency mattered more than the last few points of mAP that a two-stage detector would buy — here's the number'. Naming what you gave up, and why it was worth it, is the tell that you actually made an engineering decision rather than following a tutorial's defaults.
What distinguishes a senior-level defense of a model choice from a junior one?
