Rare Event Detection Using Disentangled Representation Learning

Abstract

This paper presents a novel method for rare event detection from an image pair with class-imbalanced datasets. A straightforward approach for event detection tasks is to train a detection network from a large-scale dataset in an end-to-end manner. However, in many applications such as building change detection on satellite images, few positive samples are available for the training. Moreover, scene image pairs contain many trivial events, such as in illumination changes or background motions. These many trivial events and the class imbalance problem lead to false alarms for rare event detection. In order to overcome these difficulties, we propose a novel method to learn disentangled representations from only low-cost negative samples. The proposed method disentangles different aspects in a pair of observations: variant and invariant factors that represent trivial events and image contents, respectively. The effectiveness of the proposed approach is verified by the quantitative evaluations on four change detection datasets, and the qualitative analysis shows that the proposed method can acquire the representations that disentangle rare events from trivial ones.

NestedVAE: Isolating Common Factors via Weak Supervision

“…Accuracy Outlier AutoEncoder [89] 0.5427 VAE [43] 0.5495 Clustering AutoEncoder [7] 0.5514 Reconstruction Prob. VAE [6] 0.5724 Adversarial VAE [60] 0.5834 Multi-Level VAE [13] 0.6072 Rare-Event VAE [32] 0.7166 NestedVAE (ours) 0.7380 Table 2. Change detection accuracy on rotated MNIST.…”

“…K-means clustering is then used to group the representation distances. Alternative results taken from [32]. The best result is shown in bold.…”

“…From a societal perspective, we may wish to achieve statistical and demographic parity such that our models do not reflect or amplify any unfairness present in our data or in society itself [65,91,35,66]. Success at these overlapping tasks has implications for a range of more specific downstream tasks including attribute transfer [80,81,33,39,94], person re-identification [8,27], change detection [32] adversarial robustness [37], and machine learning based decision processes [55,9,63].…”

“…NestedVAE is then evaluated for its usefulness at change detection using the same methodology as [32]. Images are alternately paired according to shared or not shared digit class.…”

“…The results for the change detection task are shown in Table 2. Nested VAE is evaluated against 7 other methods, some of which are specifically designed for change detection and which utilise significantly more powerful network architectures than ours [32,13,60]. It can be seen that Nest-edVAE outperforms the best alternative.…”