← Dandy Interview Insights

Dandy·Machine Learning Engineer·Technical Phone Screen·Senior

Senior
Jun 2026

Summary

Technical screen for an ML Engineer role at Dandy. One meaty computer vision / estimation problem that took up most of the time. Not a lot of small talk, they got right into it.

Questions Asked (1)

Q1

A photo has multiple people in it. The camera's focal length and each person's distance from the camera are unknown. How would you design an ML system to predict each person's actual height?

System DesignTechnical Trade-offsAlgorithms & Data Structures
Author's notes

This one has a lot of surface area and I kind of spiraled at first.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the problem as an ill-posed inverse problem where absolute height is unidentifiable from a single image without additional cues. Propose a multi-stage ML system that first estimates relative scale using human body proportions and scene context, then calibrates to absolute height using learned priors or reference objects. Discuss how to handle ambiguity through probabilistic outputs and validation strategies.

Pro tip: Acknowledge the fundamental ambiguity upfront and propose a fallback to relative height ranking when absolute prediction is impossible—this shows you understand real-world constraints and avoids overpromising. Also, mention that you'd leverage anthropometric statistics (e.g., head-to-body ratio) as a strong prior, which is often overlooked.

1. Problem Formulation and Identifiability Analysis

Clarify that without focal length or distance, absolute height is not uniquely determined from a single image. Define the goal as predicting height up to a scale factor or using learned priors to resolve ambiguity.

2. Data Collection and Annotation

Gather a dataset of images with multiple people, annotated with actual heights, camera parameters (if available), and distances. Include diverse scenes, poses, and camera setups to train robust models.

3. Feature Extraction and Relative Scale Estimation

Use a pose estimation model to detect body keypoints and compute relative proportions (e.g., head size, limb lengths). Combine with scene cues like ground plane and vanishing points to estimate relative scale between people.

4. Absolute Height Prediction with Priors

Train a model that takes relative features and outputs absolute height by learning anthropometric priors (e.g., average head-to-height ratio) and contextual cues (e.g., known object sizes). Use probabilistic regression to output a distribution.

5. Evaluation and Uncertainty Handling

Evaluate using metrics like mean absolute error and calibration of uncertainty. Implement fallback to relative height ranking when confidence is low, and discuss how to handle edge cases like children or unusual poses.

Key Points to Mention

  • Ill-posed nature of the problem: absolute height is unidentifiable without camera intrinsics or distance.
  • Use of anthropometric priors (e.g., head-to-body ratio) to resolve scale ambiguity.
  • Leveraging scene context: ground plane estimation, vanishing points, and reference objects.
  • Multi-person consistency: enforcing relative height ordering and scale consistency across people.
  • Probabilistic modeling to quantify uncertainty and provide confidence intervals.
  • Fallback strategies: predicting relative height ranking when absolute prediction is unreliable.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.