Seoul National University's DART Tracks OOD Boundaries as Environments Drift
DART tracks a drifting ID vs OOD axis at test time, cutting false positive rate by 49 points on corrupted ImageNet benchmarks.
Takeaways
- DART is a post-hoc, online OOD detector that tracks dual ID/OOD prototypes per feature layer.
- Observation: csID and csOOD stay linearly separable along a drifting axis under covariate shift.
- Uses EMA prototype updates, multi-layer fusion, and periodic flip correction against an MSP reference.
- Delivers +15.32 pp AUROC and -49.15 pp FPR@95TPR on ImageNet-C vs iNaturalist-C.
- Works with a frozen backbone, no retraining, small memory, suitable for streaming deployments.
- Code available at github.com/andylws/DART, accepted at WACV 2027.
Researchers at Seoul National University have introduced DART, a post-hoc detector that updates its in-distribution and out-of-distribution boundary as test conditions change. It runs on a frozen classifier, uses unlabeled batches, and requires no test-time gradient updates.
In-distribution (ID) samples come from the data a model was built to recognize, while out-of-distribution (OOD) samples fall outside that expected distribution. Production conditions such as rain, sensor drift, motion blur, and changing light can distort both groups at once. A detector tied to clean training statistics may then confuse environmental corruption with novelty, making its OOD scores unreliable when they are needed.
Fixed references fail under shared drift
Many post-hoc detectors compare each input with statistics captured from the training distribution. When a snowstorm or camera fault moves every feature vector, those fixed references stop describing either the familiar samples or the unknown ones.
The paper evaluates streams containing ID and OOD samples affected by the same evolving covariate shift. Covariate shift changes the appearance or measured features of an input while leaving its underlying identity intact. In this setting, an ImageNet image and an unrelated OOD image might both receive the same brightness, blur, or noise corruption.
Under that protocol, several methods built around clean training statistics fall toward an AUROC of 0.5, which indicates near-random separation across score thresholds.
A separable axis survives corruption
Layer-by-layer activation analysis showed that covariate-shifted ID and OOD samples remain linearly separable along a direction in feature space, even under severe corruption. The direction moves as conditions change, but the separation persists.