Video Anomaly Detection, and other Emerging Oddities

An open-source exploration into video anomaly detection models, evaluation benchmarks, and emerging oddities.

In Progress
Writing this openly… it is 40% done. There is a whole academic body of effort to extend it, and cite, as well as the ‘choose your own adventure’ nature of interactive reading.

I am very tired.

VAD Model Eval

Guided Tour
1 · Model Comparison
2 · Error Analysis (Spatial & Temporal)
3 · Operational Gate Validation
4 · Data Quality & Coverage Audit
5 · Incident Investigation & Triage
Document Index
Getting Started
Workflow
1 · Model Comparison
2.1 · Confusion Matrices
2.2 · Zone × Time
Dataset: full-suite-v1
250 traces · 5 models
VAD Model Evaluation Dashboard
Stage 1 of 5Model Comparison

Model Comparison

Compare the performance of all 5 iterations in our ablation study. Baseline (Frame-level) has high false alarms (14.2 FAR/100h) and low F1 (0.68). Model A4 integrates memory-guided attention, temporal smoothing, and event persistence to optimize metrics.

Diagnostic Task

Compare Baseline vs A4. Toggle candidate model A4 on and off to observe F1 score improvements and delta updates in the comparison table.

Dynamic Checkpoints
Select Candidate Model (A4) in comparison list
Keep Baseline selected to view delta comparison
Active Workspace: Ablation Study Comparison
Models:
ModelPR-AUCF1tIoUPrecisionRecallFAR/100hConfirm RateFrag. IndexLatency (s)Status
Baseline (Frame-level)
0.720
0.680
0.420
70.0%
65.0%
14.2
61%
3.40
2.8s
completed
A1: Memory-guided
+6.00.780
+6.00.740
+13.00.550
+6.076.0%
+7.072.0%
-1.412.8
+4.065%
-0.62.80
-0.42.4s
completed
A2: Temporal Smoothing
+10.00.820
+11.00.790
+22.00.640
+12.082.0%
+11.076.0%
-3.111.1
+10.071%
-1.71.70
-0.72.1s
completed
A3: Event Persistence
+13.00.850
+15.00.830
+29.00.710
+16.086.0%
+15.080.0%
-4.89.4
+15.076%
-2.31.10
-1.01.8s
completed
A4: Full Stack (Candidate)candidate
+17.00.890
+19.00.870
+36.00.780
+21.091.0%
+20.085.0%
-7.46.8
+23.084%
-2.31.05
-1.41.4s
completed