Set Features for Anomaly Detection
Authors: Niv Cohen, Issar Tzachor, Yedid Hoshen
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our method, SINBAD (Set INspection Based Aomalies Detection) is evaluated on two diverse tasks. The first task is image-level logical anomaly detection on the MVTec-LOCO datasets. We also evaluate our method on series-level time series anomaly detection. Our method outperforms more complex state-of-the-art methods while not using augmentations or training. |
| Researcher Affiliation | Academia | Niv Cohen EMAIL School of Computer Science and Engineering The Hebrew University of Jerusalem, Israel Issar Tzachor EMAIL School of Computer Science and Engineering The Hebrew University of Jerusalem, Israel Yedid Hoshen EMAIL School of Computer Science and Engineering The Hebrew University of Jerusalem, Israel |
| Pseudocode | Yes | We provide an algorithm box Alg.1 summarizing our steps. Algorithm 1 Set-based Anomaly Detection with Histogram Projections |
| Open Source Code | Yes | Our code can be found at https://github.com/Niv C/SINBAD/ |
| Open Datasets | Yes | We use the recently published MVTec-LOCO dataset Bergmann et al. (2022) to evaluate our method s ability to detect anomalies caused by unusual configurations of normal elements. Time series dataset. We evaluate on the five datasets used in Neur Tra L-AD Qiu et al. (2021): Racket Sports (RS). Epilepsy (EPSY). Naval air training and operating procedures standardization (NAT). Character trajectories (CT). Spoken Arabic Digits (SAD). |
| Dataset Splits | Yes | The anomalies in each class are divided into structural anomalies and logical anomalies. Structural anomalies feature local defects, somewhat similar to previous datasets such as Bergmann et al. (2019). Conversely, logical anomalies may violate logical conditions expected from the normal data. As one example, an anomaly may include a different number of objects than the numbers expected from a normal sample (while all the featured object types exist in the normal class; see Fig.1). Other types of logical anomalies in the dataset may include cases where distant parts of an image must correlate with one another. For instance, within the normal data, the color of one object may correlate with the length of another object. These correlations may break in an anomalous sample. |
| Hardware Specification | Yes | Computational resources. The experiments were run on a single RTX2080-GT GPU. The baseline methods were run on a single RTX2080-GT GPU. |
| Software Dependencies | No | For the whitening of image features we use the Shrunk Covariance function from the scikit-learn library Pedregosa et al. (2011) with its default parameters. For k NN density estimation we use the faiss library Johnson et al. (2019). |
| Experiment Setup | Yes | For the image experiments, we use histograms of K = 5 bins and r = 1000 projections. For the raw-pixels layer, we used a projection dimension of r = 10 and no whitening due to the low number of channels. To avoid high variance between runs, we did 32 different repetitions for the raw-pixel scoring and used the median score. We use k = 1 for the k NN density estimation. For the time series experiments: We chose a kernel size of τ = 9, r = 100 projection, K = 20 quantiles, and a maximal number of pyramid levels of L = 10, each using consecutive strides. |