Learning to Adapt to Unseen Abnormal Activities
under Weak Supervision

Jaeyoo Park* Junha Kim* Bohyung Han

ECE & ASRI, Seoul National University, Korea

Abstract

We present a meta-learning framework for weakly supervised anomaly detection in videos, where the detector learns to adapt to unseen types of abnormal activities effectively when only video-level annotations of binary labels are available. Our work is motivated by the fact that ex- isting methods suffer from poor generalization to diverse unseen exam- ples. We claim that an anomaly detector equipped with a meta-learning scheme alleviates the limitation by leading the model to an initialization point for better optimization. We evaluate the performance of our frame- work on two challenging datasets, UCF-Crime and ShanghaiTech. The experimental results demonstrate that our algorithm boosts the capa- bility to localize unseen abnormal events in a weakly supervised setting. Besides the technical contributions, we perform the annotation of missing labels in the UCF-Crime dataset and make our task evaluated effectively.


Results

Table 1: AUC (%) comparison among three different scenarios on each target subclass. All of the fine-tuning processes are the same, but the initial point of fine-tuning is different. In S algorithm, the model is fine-tuned from a random scratch model. In P scenario, the model is pretrained with Dbase before fine-tuning. Both MS and MG are meta-learning algorithms which meta-train model with Dbase to get the initial model. M*S* and M*G* correspond to "sampling" and "global" which are model selection methods of meta-training. Details of "sampling" and "global" are in Section 4.4. The bold-faced numbers correspond to the best performance for each subclass. Error while loading image

Paper

ACCV 2020 paper. (pdf, 2.4MB)

Citation

Jaeyoo Park, Junha Kim, and Bohyoung Han. Learning to Adapt to Unseen Abnormal Activities under Weak Supervision, Asian Conference on Computer Vision (ACCV) 2020
Bibtex


Code

dataset, (600MB) including all features and annotations.
github