Boundary-Aware Behavior Analytics. A self-hosted NVR that doesn't just record — it watches, recognizes and remembers. Entirely on your own GPU.
Full-frame transformer detection
RT-DETRv2, D-FINE and DEIMv2 run on every frame — no motion gating. A person standing still never disappears, and panoramic cameras get automatic tiled inference.
Real re-identification
DINOv2 body embeddings cluster the same person across cameras and across days. Name someone once; every past and future sighting links to them.
Face recognition
YuNet/SCRFD detection, ArcFace alignment, AuraFace embeddings — with photo enrollment and an operator review queue. Commercial-clean model licenses throughout.
Scene states
Few-shot visual classifiers watch regions of the frame: gate open or closed, garage door up or down, lights on or off. Train from a handful of crops.
Zones with SAM2
Draw polygons — or click an object and let SAM2 segment it pixel-perfectly. Enter, exit and dwell events with per-zone rules.
Visual search
Drop in a photo and find matching people across your entire history via vector search.
24/7 recording
Continuous segments with codec passthrough, a scrubbable timeline, clip extraction and tiered retention.
Notifications
SMTP, Slack, Telegram and webhooks with a rules engine and a full delivery audit trail.
See it work
Run it yourself
One wizard, prebuilt images from GHCR, models downloaded automatically. NVIDIA is the reference target (an RTX 3060 carries a full home), Intel Arc/iGPU is supported, and plain CPU works for evaluation.
git clone https://github.com/Bacinac/baba.git
cd baba
./install.shCurrent release: v0.1.253 ·
images public on ghcr.io/bacinac/baba/* ·
public repository opening shortly.