TACnet: Weakly Supervised Video Anomaly Detection

Designed and implemented a pipeline that trained on one normal or anomalous label per video while producing anomaly scores for individual temporal segments.

Problem

Videos had different lengths, anomalous events occupied only a small fraction of their frames, and frame-level training labels were too expensive to assume.

Implementation

Sampled every video into 32 segments of 16 frames, used a Sports-1M-pretrained C3D backbone, and added a 1D temporal convolution to capture longer-range context while reducing the feature dimension.

Reasoning

A uniform segment sequence made variable-length videos batchable, while retaining the video backbone inside the training path avoided treating precomputed clip averages as the final temporal representation.

Outcome

Created one end-to-end representation for video-level classification and segment-level localization across UCF-Crime and ShanghaiTech.

Combined gated temporal attention with two-cluster K-means to infer which segments inside an anomalous video represented the event.

Problem

A video-level label identifies an anomalous bag but does not reveal which of its segments contain the anomaly.

Implementation

Learned attention weights for the bag prediction, concatenated each segment feature with its attention weight, clustered the segments with cosine distance, and treated the cluster with the higher mean segment score as the anomaly candidate.

Reasoning

Attention exposed each segment's contribution to the bag decision, while clustering converted that weak signal into a within-video separation without requiring frame labels. Cosine distance was more stable than Euclidean distance in the experiments.

Outcome

Removing the clustering path produced the largest top-down ablation drop, 1.97 percentage points on UCF-Crime.

Proposed an entropy smoothness loss and evaluated the complete model through quantitative comparisons, qualitative traces, ablations, and clustering-distance experiments.

Problem

Penalizing differences between adjacent segment scores can blur the boundary between normal and anomalous behavior.

Implementation

Quantized segment scores into 20 intervals and minimized their distribution entropy, then analyzed component removals, attention weights, prediction curves, and feature clusters.

Reasoning

Concentrating predictions into fewer score ranges encouraged stable, separable outputs without assuming neighboring segments must always receive similar scores.

Outcome

The model achieved 84.75% frame-level AUC on UCF-Crime and 93.89% on ShanghaiTech in the paper's evaluation.