Bulwark: driver drowsiness and distraction detection models
Research models that detect signs of driver drowsiness and distraction. Non-commercial weights. Commercially licensed models: inferensys.com/contact
Video Classification • 93.6M • Updated • 5 • 1Note Bulwark-B1, the base video model: Meta's V-JEPA 2.1 encoder with LoRA adapters. From a 2-second clip of a camera facing the driver, it detects eyes closed, PERCLOS, blinks, yawns, phone use, eating and drinking, and the driver's activity. Research preview: indicative results on 3 test people per dataset; not certified or tested against the EU General Safety Regulation or Euro NCAP. Weights: CC BY-NC 4.0, non-commercial research only. Commercially licensed models: https://inferensys.com/contact
Inferensys/bulwark-e1-driver-monitoring-onnx
Video Classification • Updated • 2 • 1Note Bulwark-E1, the CPU model: an EfficientViT-B1 student distilled from Bulwark-B1 and exported to ONNX, as a 16-frame video graph and a 1-frame image graph. E1 v0 is a research preview that scores well below B1: use it to try CPU deployment, and B1 for research. Licence and caveats as B1: CC BY-NC 4.0, non-commercial research only. Commercially licensed models: https://inferensys.com/contact
DMD: A Large-Scale Multi-Modal Driver Monitoring Dataset for Attention and Alertness Analysis
Paper • 2008.12085 • PublishedNote DMD, the Driver Monitoring Dataset (Vicomtech): one of Bulwark's four training datasets and its source of blink and yawn ground truth. Academic use only (CC BY-NC-ND 4.0).
V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning
Paper • 2603.14482 • Published • 40Note V-JEPA 2.1 (Meta): the self-supervised video encoder that Bulwark-B1 builds on. B1 uses Meta's ViT-B/16 checkpoint with LoRA adapters. Meta publishes it in its V-JEPA 2 repository, which is under the MIT License and states no separate licence for its checkpoints.
EfficientViT: Lightweight Multi-Scale Attention for On-Device Semantic Segmentation
Paper • 2205.14756 • Published • 1Note EfficientViT (MIT Han Lab): the backbone of Bulwark-E1, initialised from timm's ImageNet-1k weights (timm/efficientvit_b1.r224_in1k, Apache-2.0).
Driver Distraction Identification with an Ensemble of Convolutional Neural Networks
Paper • 1901.09097 • PublishedNote AUC Distracted Driver dataset paper: one of Bulwark's four training datasets (non-commercial licence agreement; use limited to distracted-driver detection research).