Masked multi-prediction for multi-aspect anomaly detection
Authors: Yassine Naji, Romaric Audigier, Aleksandr Setkov, Angelique Loesch, Michèle Gouiffès
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experiments conducted on several benchmarks show the effectiveness of the proposed approach. Preliminary experiments to support our analysis have been carried out on the following datasets: Synthetic roads dataset, MNIST. Section 5 Experimental study. This section presents the results (Table 3) of our method on the benchmarks UCSDped2, Avenue, Shanghai Tech. Since we focus on object-level anomalies and for a fair comparison with state-of-the-art objectcentric methods, we use the same baseline object-detector (Yolov3). |
| Researcher Affiliation | Academia | Yassine Naji EMAIL Université Paris-Saclay, CEA, List, 91120, Palaiseau, France Université Paris-Saclay, CNRS, LISN, 91400, Orsay, France; Romaric Audigier EMAIL Université Paris-Saclay, CEA, List, 91120, Palaiseau, France; Aleksandr Setkov EMAIL Université Paris-Saclay, CEA, List, 91120, Palaiseau, France; Angélique Loesch EMAIL Université Paris-Saclay, CEA, List, 91120, Palaiseau, France; Michèle Gouiffès EMAIL Université Paris-Saclay, CNRS, LISN, 91400, Orsay, France |
| Pseudocode | Yes | In order to facilitate the reproducibility of our approach, we provide the pseudo-code 1 for training MMP-AMS. |
| Open Source Code | No | The paper provides pseudo-code in Appendix B.1, but does not explicitly state that the source code for the methodology described is publicly released or provide a link to a repository. It mentions using implementations of third-party tools (MMDetection, Flow Net2) but not its own code. |
| Open Datasets | Yes | Preliminary experiments to support our analysis have been carried out on the following datasets: Synthetic roads dataset, MNIST (Le Cun et al. (2010)). We perform experiments on the most commonly used datasets for the one-class and object-centric scenario. UCSDped2 (Mahadevan et al. (2010)). Shanghai Tech (Luo et al. (2017)). CUHK Avenue (Lu et al. (2013)). |
| Dataset Splits | Yes | MNIST (Le Cun et al. (2010)): this dataset is used to show the importance of masking. It contains handwritten digits from 0 to 9. As this dataset was not designed for the one-class setting, it is adapted by considering a particular class as normal in each case. We use the training-testing split of the dataset and perform training only on samples from the class considered normal. |
| Hardware Specification | Yes | Regarding the inference time, our model processes a batch of objects in a frame taken from Avenue in 18ms on a single Nvidia Titan-X GPU. |
| Software Dependencies | No | The paper mentions using Yolov3 (Redmon & Farhadi (2018)) with MMDetection (Chen et al. (2019)) and Flow Net2 (Reda et al. (2017)), but it does not specify version numbers for these software components or any other libraries/frameworks. |
| Experiment Setup | Yes | We train the network for 150 epochs for UCSDped2 and Avenue and for 400 epochs for Shanghai Tech using Adam optimizer with a learning rate of 10 3 with a batch size of 640 for the biggest dataset Shanghai Tech and 64 for UCSDped2 and Avenue. For the mask M, we remove 50% of pixels using a grid of 4x4 pixels. Regarding the distance used for anomaly scoring in Section 4, we use the L1 distance for RGB and optical flow, as well as the cross entropy loss for class probabilities. We set (γ, w I, w F , w C, w BX, w BY , w BH, w BW , λ) = (0, 1, 1, 1, 1, 1, 1, 1, 0.1) for UCSDped2 and Shanghai Tech and (1, 1, 0.1, 0.1, 1, 1, 1, 1, 0.1) for Avenue. For the non-participation loss, we select predictors which have a participation below δ = 5%. |