Add PLF postprocessor

[maryhadalittlecode]

Adds PLF, a post-hoc OOD detector that applies a piecewise linear shaping function to penultimate-layer features.

Paper: "An Information-Theoretic Framework for Feature Construction in Out-of-Distribution Detection": https://arxiv.org/abs/2506.14194

Leaderboard PR:
The corresponding leaderboard entries were submitted separately in zjysteven#7: zjysteven#7.

Our git repo: https://github.com/maryhadalittlecode/PLF-OpenOOD
PLF’s hyperparameters are tuned using Bayesian optimization (BO); please refer to our repo for the complete tuning procedure, as the grid search range in configs/postprocessors/plf.yml included here is provided only as an example.

Changes

• openood/postprocessors/plf_postprocessor.py — PLF postprocessor implementation
• configs/postprocessors/plf.yml — default configuration and hyperparameters
• openood/networks/plf_net.py — network wrapper for applying PLF to penultimate features
• openood/postprocessors/__init__.py, openood/postprocessors/utils.py, and openood/evaluation_api/postprocessor.py — registration under the 'plf' key
• openood/evaluation_api/evaluator.py — integration of the PLF network wrapper into the evaluation pipeline
• scripts/ood/plf/ — evaluation scripts for CIFAR-10, CIFAR-100, ImageNet-200, and ImageNet-1K

Results

CIFAR-10 (ResNet-18)

Category	Method	Near-OOD	Far-OOD	Avg-OOD	Ranking
Feature Shaping	PLF (Ours)	89.47 ± 0.89	92.36 ± 1.09	91.40 ± 1.48	3

CIFAR-100 (ResNet-18)

Category	Method	Near-OOD	Far-OOD	Avg-OOD	Ranking
Feature Shaping	PLF (Ours)	80.75 ± 0.19	80.53 ± 1.38	80.60 ± 1.24	3

ImageNet-200 (ResNet-18)

Category	Method	Near-OOD	Far-OOD	Avg-OOD	Ranking
Feature Shaping	PLF (Ours)	83.28 ± 1.11	93.82 ± 0.85	89.60 ± 0.97	2

ImageNet-1K (ResNet-50)

Category	Method	Near-OOD	Far-OOD	Avg-OOD	Ranking
Feature Shaping	PLF (Ours)	78.15	96.24	89.00	1