The authors apply post-hoc OOD detectors to ImageNet and CIFAR-100 backbones.
OOD detection gains mostly reflect dataset identity, not novelty
By testing on held-out OOD datasets rather than held-out samples, the authors show that most reported improvements vanish.
Academic
Donghoon Lee · Shinjin Kang
Hongik University
Research Digest··2 min read
Lee and Kang measured how much of the reported out-of-distribution detection gain survives when the detector is tested on an OOD dataset it has never seen—not just on held-out samples of the same dataset.
Why this paper
From Hongik University
In one line
Most reported OOD detection gain is inflated by dataset identity, not novelty.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§