Hypercone Assisted Contour Generation for Out-of-Distribution Detection
Annita Vapsi, Andr\'es Mu\~noz, Nancy Thomas, Keshav Ramani, Daniel, Borrajo
TL;DR
HAC_k-OOD is a novel out-of-distribution detection method that constructs hypercones to adaptively approximate in-distribution data contours, achieving state-of-the-art results without explicit OOD training.
Contribution
It introduces hypercone-based contour approximation for OOD detection that adapts to data distribution without assumptions, improving detection performance.
Findings
Achieves state-of-the-art FPR@95 and AUROC on CIFAR-100
Effective for both Near-OOD and Far-OOD detection
Does not require explicit OOD training
Abstract
Recent advances in the field of out-of-distribution (OOD) detection have placed great emphasis on learning better representations suited to this task. While there are distance-based approaches, distributional awareness has seldom been exploited for better performance. We present HAC-OOD, a novel OOD detection method that makes no distributional assumption about the data, but automatically adapts to its distribution. Specifically, HAC-OOD constructs a set of hypercones by maximizing the angular distance to neighbors in a given data-point's vicinity to approximate the contour within which in-distribution (ID) data-points lie. Experimental results show state-of-the-art FPR@95 and AUROC performance on Near-OOD detection and on Far-OOD detection on the challenging CIFAR-100 benchmark without explicitly training for OOD performance.
Peer Reviews
Decision·Submitted to ICLR 2025
HACk-OOD introduces a unique method using hypercone projections to delineate class contours, avoiding traditional Gaussian distribution assumptions and offering greater flexibility in complex feature spaces. The method achieves competitive, often superior, results on challenging datasets like CIFAR-100, demonstrating strong performance in both near and far OOD detection scenarios.
1. Experiments are limited to CIFAR-based datasets, testing on a large-scale dataset like Imagenet would better validate the method’s scalability. Also evaluating HACk-OOD on the OpenOOD benchmark would provide a clearer comparison to recent methods. I would consider rating this paper higher if Imagenet results were provided. 2. Missing comparisons with some of the latest post-hoc OOD methods, such as ASH and SCALE. Including these would offer a more comprehensive assessment of its relative per
1. The method takes an interesting approach to distance-based OOD detection by relaxing the distributional assumptions and, unlike naive KNN, still leveraging nearby training data statistics to construct class contours. To the best of my knowledge, the use of hypercones for this purpose is novel and appears well-motivated. 2. The authors present their method clearly, making the paper easy to follow and understand. 3. Although the method involves hyperparameter k, the authors provide a practica
1. The computational complexity is a concern for this method. Since the computation appears to increase with the size of the training dataset, it’s unclear if this approach would be feasible for large-scale, real-world applications. Although the authors state that the method is computationally efficient and support this with inference time per sample, I encourage them to provide a more detailed discussion on this aspect. For instance, what is the time required to construct the hypercones? A comp
1. The paper is well-written and easy to follow. The authors provide a good summary of the different approaches to OOD, a background section on hypercones, and a clear and precise method description. 2. Relevance and novelty of the method: the algorithm doesn't require assumptions about the data distribution and can model complex embedding spaces since it draws multiple hypercones per class and since it defines per-hypercone decision boundaries. 3. The authors discuss some limitations of the m
1. Limited evaluation: the method is only evaluated on models pre-trained on the *quite simple* CIFAR datasets and not on more complex datasets such as the ImageNet-200 or ImageNet-1k OOD benchmarks.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsIndustrial Vision Systems and Defect Detection · Advanced Measurement and Detection Methods · Flow Measurement and Analysis
MethodsSparse Evolutionary Training