An arXiv preprint reports that a video model trained on coarse, clip-level risk supervision can also identify the moments and tracked entities on which its prediction depends. On 179 egocentric driving scenarios, the model recorded the lowest reported average prediction error, 0.575, and the highest reported Spearman correlation, 0.655, among the listed methods. Its root mean square error was 0.724, and its pairwise ranking accuracy was 0.748.
From a clip score to a finer map
CoRE's central move is to use prediction changes as weak supervision. It freezes a video-level predictor, intervenes on candidate temporal regions or entity tracks, measures the change in the prediction and converts that effect into graded targets. A student model then learns to predict support directly, so interventions are not needed at inference. The learned support is defined as the dependence of the coarse prediction on the candidate that was perturbed, not as proof that the candidate physically causes risk.
On RISEE, the temporal-support test reported a selected prediction drop of 0.320 and a gain over random interventions of 0.245. The correlation between intervention effects and learned support was 0.542. The reported top-percentile score was 0.888, and the positive rate was 0.989.
At the entity level, CoRE reported a selected entity-effect drop of 0.243 and a gain over random of 0.117.
RISEE contained 179 egocentric driving scenarios with aggregated human perceived-risk ratings, and evaluation used five-fold evaluation at the scenario level. Because temporal risk intervals and risk-entity labels were unavailable, support was tested through held-out candidate interventions rather than human localization ground truth. Fine-grained temporal or entity annotations were not used for training.
Testing anomaly localization
The approach was also evaluated in a weakly supervised DoTA setup. Training used 2,420 anomalous and 3,234 normal videos, while validation used 269 anomalous and 358 normal videos. The released test split contained 1,140 videos.
With ResNet-50 features, the reported frame AUC was 0.735, F1 at the 0.5 setting was 0.374, and best temporal intersection-over-union, or tIoU, was 0.440, compared with 0.642, 0.237 and 0.370, respectively, in the comparison. With CLIP ViT-B/32 features, CoRE reported AUC 0.744, average precision, or AP, of 0.514, F1 at 0.5 of 0.364 and best tIoU of 0.429.
On UCF-Crime, CoRE achieved a frame AUC of 85.68%, compared with 84.30% for a controlled reproduction of RTFM and 82.79% for a controlled reproduction of MGFN. The split used 1,610 training videos and 290 test videos; 161 training videos were held out for validation, leaving 1,449 for optimization.
What changed when parts were removed
A fixed-configuration DoTA ablation compared the full model with versions without effect supervision or soft targets. The full configuration reported AUC, AP, F1 and tIoU of 0.744, 0.514, 0.364 and 0.429. Without effect supervision, the figures were 0.419, 0.264, 0.139 and 0.329. Without soft targets, they were 0.453, 0.283, 0.133 and 0.326.
Without multi-scale candidates, the reported AUC, AP, F1 and tIoU values were 0.504, 0.314, 0.040 and 0.098. Removing the bag objective produced a mixed result: AUC and AP were 0.754 and 0.518, while F1 and tIoU were 0.309 and 0.335.
A useful result with a clear boundary
The paper also compared local-mean and blur interventions. For temporal candidates, the effects had a reported correlation of 0.7344, top-one agreement of 0.6089 and top-three Jaccard overlap of 0.6765. For object candidates, the corresponding figures were 0.7356, 0.8333 and 0.8512.
CoRE's support scores measure how much the learned coarse prediction depends on a candidate under a specified intervention; they do not establish physical causality. On RISEE, that means the support output was evaluated through held-out interventions rather than against human time-and-entity localization labels. The work is identified as arXiv:2608.25344v1, dated 26 Aug 2026.
Paper data and sources
Original title: CoRE: Weakly Supervised Coarse-to-Fine Risk Evidence Learning in Driving Videos
Authors: Kaiser Hamid, Can Cui, Nade Liang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text