Detecting anomalous behaviour in surveillance video is a high-impact application of machine learning, but supervision is available only at the video level: a set of temporal segments carries a single anomaly or normal label, with no per-segment ground truth. This is the Multiple Instance Learning setting. On the UCF-Crime benchmark, published weakly-supervised methods have converged to a narrow band of Receiver Operating Characteristic Area Under the Curve between 0.85 and 0.87 over the past three years, despite the adoption of progressively larger Vision Transformer backbones. We investigate whether this ceiling reflects a limit on feature quality or a limit on prediction diversity. We introduce HRM-Crime, a compact hierarchical score model of approximately 270 thousand trainable parameters that couples windowed self-attention for local temporal structure with dilated one-dimensional convolutions for long-range temporal context. Segment features are pre-extracted with a frozen ResNet-50 for appearance and a frozen R3D-18 for motion. On this base we evaluate two orthogonal interventions: aggregating each model’s segment scores by the mean of its K highest values rather than the maximum, and fusing three models trained on different feature subsets through rank normalisation. Across 98 training runs a single model attains a mean Area Under the Curve of 0.8934 with standard deviation 0.0087. The three-stream ensemble attains 0.9300, an improvement of 0.0213 over the strongest single model, with a bootstrap 95 percent confidence interval of 0.0079 to 0.0360 and a one-sided p-value of 0.0008. Five negative results are documented. We note that the reported gain confounds architecture, feature representation and ensembling, and we identify the experiment required to separate them.

Automated detection of anomalous behaviour in surveillance video is a high-impact application of machine learning, but training data is available only at the video level: a bag of temporal segments is labelled anomaly or normal with no per-segment ground truth. Prior weakly-supervised methods on the UCF-Crime benchmark have plateaued around 85–87% ROC-AUC using ever larger Vision Transformer backbones and increasingly elaborate score modules. We take the opposite approach. This paper introduces HRM-Crime, a compact hierarchical model (approximately 270,000 parameters) that combines Swin-style windowed self-attention for local temporal reasoning with dilated 1-D convolutions (dilations 1 → 2 → 4) for global temporal context. Features are pre-extracted as ResNet-50 spatial vectors (2048-D) and R3D-18 temporal vectors (512-D), L2-normalised and concatenated. Beyond the architecture, two orthogonal contributions further improve performance: per-model top-K score aggregation with K tuned per feature stream, and a three-stream rank-normalised ensemble over combined, spatial-only, and temporal-only variants. On UCF-Crime (290 test videos) the ensemble reaches ROC-AUC = 0.9303, PR-AUC = 0.9192, EER = 0.128, detection F1 = 0.870, and IoU (Jaccard) = 0.771, outperforming the strongest published ViT-B weakly-supervised baseline (UR-DMU, 0.8697) by +6.0 absolute AUC points while using two to three orders of magnitude fewer parameters and less compute at inference. Five negative results are also documented for reproducibility.