← all challenges

// challenge brief · hack a ton 2026

How Far?

Given a cropped object and its class, estimate how many meters away it is, accurately, from one image.

// Proposed by
Monsson
// Industry
Computer vision / robotics (Unitree Go2 Edu)
// Difficulty
🟡 Intermediate

★ Own jury, own €1,000 prize. Pick any one of the three to compete.

// the problem

Knowing how far away something is normally needs stereo or a depth sensor. A single camera doesn't measure distance directly. But combine a depth network, what you know about an object's real size, and the camera's geometry, and you can estimate it surprisingly well. This is the cleanest, best-scoped ML problem on the Monsson track: one frame in, one number out, one clear metric.

// your mission

Given a crop from a frame and the object class in it, estimate the distance in meters from the robot's camera to the object, with mean absolute error below the threshold.

// how it works

  1. 1Perceive: read the crop and its class label.
  2. 2Estimate: combine neural depth, object size priors, and camera geometry (intrinsics provided) into a distance.
  3. 3Report: output the distance, and (bonus) a confidence interval.

Target accuracy: MAE < 15% on the hidden test set.

// what we provide

  • A set of images with common objects (chairs, people, doors, pots, balls) at measured distances.
  • The Go2 camera intrinsics.
  • A hidden test set for scoring.

// suggested approach

  • A pretrained monocular depth model (Depth Anything / MiDaS) for a relative depth cue.
  • Pinhole geometry: known real-world object height + pixel height + focal length → metric distance. Fuse the two.
  • Floor version: geometry-from-size alone gets you a real baseline. Layer the depth network on top to improve it.

// how you'll be judged

  • MAE on the hidden test set.
  • Behavior at extreme distances (out-of-distribution).
  • Quality of the uncertainty estimate.
  • Methodological rigor in your reporting.

// stretch goals

  • Calibrated confidence intervals, not just point estimates.
  • Graceful degradation when the object class is unusual or the crop is poor.

// deliverables

  • Git repo: code + README + run instructions.
  • Demo: notebook or script on the provided test set.
  • 5-minute pitch: architecture, decisions, trade-offs.