Start by clarifying the problem constraints and available data, then propose a two-stage approach: first estimate depth or relative scale using monocular cues, then convert to real-world height using a reference object or learned priors. Discuss trade-offs between end-to-end learning and modular pipelines, and emphasize the need for a calibration strategy.
Pro tip: Acknowledge that absolute height estimation from a single image is ill-posed without a reference; propose using known object sizes (e.g., doors, furniture) or human pose priors to resolve scale ambiguity, and suggest a fallback to relative height ranking if absolute values are not critical.
Ask about the use case: is absolute height needed or relative ranking? What data is available (single image, multiple images, metadata)? What accuracy is acceptable? This shapes the entire design.
Decide between using a reference object of known size, leveraging human pose and anthropometric priors, or training a model to predict depth and focal length jointly. Discuss pros and cons of each.
Outline a modular pipeline: person detection, pose estimation, depth estimation, and height regression. Alternatively, propose an end-to-end model that takes an image and outputs per-person height, but note data requirements.
Discuss how to obtain training data: synthetic data with known heights, or real images with ground-truth heights. Mention data augmentation and the challenge of domain shift.
Define evaluation metrics (MAE, RMSE, rank correlation). Propose validation on a held-out set and discuss potential failure cases (occlusion, unusual poses, extreme distances).
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.