Awesome

Code for the NINCO OOD dataset introduced in

In or Out? Fixing ImageNet Out-of-Distribution Detection Evaluation

Julian Bitterwolf*, Maximilian Müller*, Matthias Hein

University of Tübingen
Tübingen AI Center

*Equal contribution

Paper: https://arxiv.org/abs/2306.00826
NINCO dataset direct download: https://zenodo.org/record/8013288/files/NINCO_all.tar.gz?download=1
NINCO Zenodo page: https://zenodo.org/record/8013288

Paper abstract

Out-of-distribution (OOD) detection is the problem of identifying inputs which are unrelated to the in-distribution task. The OOD detection performance when the in-distribution (ID) is ImageNet-1K is commonly being tested on a small range of test OOD datasets. We find that most of the currently used test OOD datasets have severe issues, in some cases more than 50% of the dataset contains objects belonging to one of the ID classes. These erroneous samples heavily distort the evaluation of OOD detectors. As a solution, we introduce with NINCO a novel test OOD dataset, each sample checked to be ID free, which with its fine-grained range of OOD classes allows for a detailed analysis of an OOD detector’s strengths and failure modes, particularly when paired with a number of synthetic “OOD unit-tests”. We provide detailed evaluations across a large set of architectures and OOD detection methods on NINCO and the unit-tests, revealing new insights about model weaknesses and the effects of pretraining on OOD detection performance.

The NINCO dataset

The NINCO (No ImageNet Class Objects) dataset consists of 64 OOD classes with a total of 5879 samples. The OOD classes were selected to have no categorical overlap with any classes of ImageNet-1K. Then we inspected each sample individually to not contain ID objects.

Examples of individual sorting decisions:

Top: Samples that were excluded due to overlap with ID classes. Bottom: Samples from the same OOD class that were included in the cleaned datasets.

Examples images from some of the most difficult NINCO OOD classes

Examples from NINCO OOD classes (first row) and from the ImageNet-1K class (second row) which the OOD class is most frequently confused for.

Cleaned Popular Datasets and OOD Unit Tests

Besides NINCO, we provide (in the same .tar.gz file) truly OOD versions of 11 popular OOD datasets with in total 2715 OOD samples.

Further included are 17 OOD unit-tests, for which 400 samples each are also included in the download. You can replicate or modify the unit-tests by following create_unit_test_data.ipynb or using noise_functions.py.

Using this repository:

NINCO dataset utilities and example notebooks for replicating the results of the paper

This repository contains the evaluation pipeline for obtaining the results in the paper. It was tested with the python package versions that are installed via pip install -r versioned_requirements.txt, though pip install -r versioned_requirements.txt should work as well (potentially using timm==0.6.11 for the same model checkpoints).

Five jupyter notebooks can in the current version be used as tutorials for the repo's functionalities:

plots_in_or_out_paper.ipynb opens the file model_outputs/df.pkl which contains most evaluation results shown in the paper. It can be modified in order to inspect individual evaluation numbers, e.g. the performance of a specific detector on a specific OOD class. These results can be accessed without downloading data or running models.
analyze_evaluations.ipynb shows example methods for examining results of such evaluations are shown in analyze_evaluations.ipynb.

NINCO dataset download

To evaluate models and view the NINCO images, please download and extract the dataset contained in this tar.gz file.

Google Drive mirror.

Run evaluations

Please edit data/paths_config.py to set ninco_folder to the folder where the downloaded datasets have been extracted (containing the folders NINCO_OOD_classes, NINCO_OOD_unit_tests and NINCO_popular_datasets_subsamples). Also, set repo_path to the absolute path of this repository.

Models can be evaluated by running evaluate.py.

For example, the ViT-B-384-21k with MSP can be evaluated on the NINCO OOD classes with:

python evaluate.py --model_name vit_base_patch16_384_21kpre --dataset NINCO --method MSP

For other datasets, use --dataset NINCO_popular_datasets_subsamples, --dataset NINCO_OOD_unit_tests, or --dataset file_list.csv where file_list.csv has a header row, one column containing the image locations and an optional second column with labels, e.g. as in data/NINCO_OOD_classes.csv.

Specifying --method all evaluates all methods, but takes considerable time since on a model's first run it requires a forward pass over the whole ImageNet-1K train set.

Example methods for examining results of such evaluations are shown in analyze_evaluations.ipynb.

Citations:

When using NINCO, please consider citing (besides this paper) the following data sources that were used to create NINCO:

Hendrycks et al.: ”Scaling out-of-distribution detection for real-world settings”, ICML, 2022.  
Bossard et al.: ”Food-101 – mining discriminative components with random forests”, ECCV 2014.  
Zhou et al.: ”Places: A 10 million image database for scene recognition”, IEEE PAMI 2017.  
Huang et al.: ”Mos: Towards scaling out-of-distribution detection for large semantic space”, CVPR 2021.  
Li et al.: ”Caltech 101 (1.0)”, 2022.
Ismail et al.: ”MYNursingHome: A fully-labelled image dataset for indoor object classification.”, Data in Brief (V. 32) 2020.
The iNaturalist project: https://www.inaturalist.org/

When using NINCO_popular_datasets_subsamples, additionally to the above, please consider citing:

Cimpoi et al.: ”Describing textures in the wild”, CVPR 2014.  
Hendrycks et al.: ”Natural adversarial examples”, CVPR 2021.  
Wang et al.: ”Vim: Out-of-distribution with virtual-logit matching”, CVPR 2022.  
Bendale et al.: ”Towards Open Set Deep Networks”, CVPR 2016.  
Vaze et al.: ”Open-set Recognition: a Good Closed-set Classifier is All You Need?”, ICLR 2022.  
Wang et al.: ”Partial and Asymmetric Contrastive Learning for Out-of-Distribution Detection in Long-Tailed Recognition.” ICML, 2022.  
Galil et al.: “A framework for benchmarking Class-out-of-distribution detection and its application to ImageNet”, ICLR 2023.

For citing our paper, we would appreciate using the following bibtex entry (this will be updated once the ICML 2023 proceedings are public):

@inproceedings{
bitterwolf2023ninco,
title={In or Out? Fixing ImageNet Out-of-Distribution Detection Evaluation},
author={Julian Bitterwolf and Maximilian Mueller and Matthias Hein},
booktitle={ICML},
year={2023},
url={https://proceedings.mlr.press/v202/bitterwolf23a.html}
}