Release date
24th October 2025
Changes
Added 1103 Active Learning images to the dataset. These samples were collected from our field devices based on model uncertainty signals (e.g., low confidence scores and ambiguous predictions) from previous model versions. They are intended to improve overall model performance.
Added 687 Reactive Learning images to the dataset. These are false positives and false negatives identified by customers and internal algorithms. Adding them follows a human-in-the-loop approach to address recurrent field issues and improve model robustness.
Added 976 images of people in casual clothing to the dataset to improve general person detection.
Model evaluation
Evaluation metrics
Model 4.26.0 achieves object detection metrics that are slightly better than 4.25.0 on the same test set, with slightly improved AP@0.5 and F0.45 scores across almost all classes (person, safety cone, and vehicle).
AI Model Version | AP@0.5 (person) | AP@0.5 (safety cone) | AP@0.5 (vehicle) | F0.45 (person) | F0.45 (safety cone) | F0.45 (vehicle) |
4.26.0 | 0.927 | 0.842 | 0.819 | 0.947 | 0.911 | 0.888 |
4.25.0 | 0.926 | 0.851 | 0.818 | 0.945 | 0.911 | 0.886 |
QA metrics
Both the current and baseline models are evaluated on 817 videos (over 320,000 labeled frames) to simulate real-world performance for the person class. Classification accuracy measures the proportion of frames that were correctly identified as containing people or not containing people.
| 4.25.0 | 4.26.0 | %change |
False positives occurrences | 45 | 24 | -46.7% |
Classification accuracy | 0.9691 | 0.9694 | +0.03% |