DMODE: Differential Monocular Object Distance Estimation Without Class-Specific Information
DMODE enables object distance estimation from a single camera without relying on object-class priors, making the approach robust to out-of-distribution inputs - a property central to safety-critical systems. The class-agnostic design reflects the principle that reliable AI must not fail silently on novel inputs it was never trained on. Published at RoMoCo 2024, this work informs how Fact AI Lab designs AI systems that fail gracefully rather than confidently incorrectly.
Accurate distance estimation from a single monocular camera is a fundamental challenge in mobile robotics - particularly when objects of interest are novel or lack class-specific training data.
The Problem with Class-Prior Methods
Most monocular depth estimation approaches resolve scale ambiguity using class priors: they assume objects of known categories (cars, pedestrians, chairs) have predictable physical dimensions. This works when the deployment environment matches the training distribution. It fails silently when objects are novel, occluded in unusual ways, or appear at unexpected scales.
Silent failure - where the system produces a confident estimate that happens to be wrong - is the most dangerous failure mode in safety-critical applications.
DMODE's Approach
DMODE (Differential Monocular Object Distance Estimation) exploits the relationship between a detected bounding box and the corresponding region in a monocular depth map, using differential analysis of depth features rather than class-based priors.
The key insight: the spatial gradient of depth values within and around a bounding box encodes reliable distance information that does not depend on knowing what the object is. By analyzing how depth values change across the bounding box boundary rather than their absolute values, DMODE produces estimates that generalize across object categories.
Properties:
- Class-agnostic: no object category labels required at inference
- Single-camera: standard monocular RGB; no stereo or depth sensor required
- Plug-in: integrates with any object detector that produces bounding boxes
Results
Evaluated on standard robotics benchmarks and a novel test set containing out-of-distribution objects (industrial equipment, construction machinery) not present in training data. DMODE maintains consistent error rates on novel objects; class-prior baselines show significantly degraded performance, confirming the failure mode the design targets.