← Dhi Labs E3 · EDGE-SCENE-GRAPHS

Neuro-symbolic scene graphs (edge-scene-graphs)

Turns detector output into queryable relations plus a symbolic rule engine, with plain-language explanations, on a zero-dependency core that runs on Jetson-class devices as-is.

Closed source, inquire

What it does

Edge-scene-graphs sits between any detector and an alerting system: geometric spatial and temporal predicate extractors (near, overlaps, in a zone, loitering, crossed a line, and more), an interval-compressed scene graph, and a declarative rule engine (joins, sequences, absence, cooldowns) that produces grounded, explainable alerts plus an ad-hoc relation query language.

The core is Python-standard-library only, no numpy, no torch, so it runs unmodified on Jetson-class hardware next to a TensorRT detector.

It also ships a full learned-relation-head path: a public scene-graph benchmark data converter, training tooling, a trained torch MLP checkpoint, and a Recall-at-K evaluation harness.

Evidence

Synthetic proof run

A scripted 60-second, 10-frames-per-second scenario processed 600 frames into 12,454 triples and 54 intervals, firing 10 alerts across 3 rule types, in 0.1357 seconds (about 4,420 frames per second processed).

Measured, local hardware (MacBook Apple Silicon, single core, pure Python, 12 objects/frame)

Mean 1.981 ms per frame, p95 2.233 ms, capacity about 504.7 frames per second.

Real hardware (Orin Nano Super, 6-core ARM, JetPack R36.4.3, live production node)
Objects/frameMean msp50 msp95 msfps capacity
62.2822.4432.663438.3
129.9269.90513.435100.7
2025.3525.27331.55239.4
Real data (public scene-graph benchmark, 50-predicate val split, 9,374 images, 61,672 triples) and real hardware
R@20mR@20R@50mR@50
Trained head0.5000.1590.5960.200
Class-pair frequency baseline0.2200.0610.3900.118
Geometric baseline (vocab mismatch)0.0070.014--

Trained on an RTX 5070 Ti, 30 epochs, about 2 minutes. The same head, exported to ONNX and built into a TensorRT 10.3 FP16 engine, runs on an Orin Nano Super at 0.117 ms mean and 0.122 ms p95 per batch-64 inference (about 9,404 queries per second), a 334 KB engine file, 0.078 MiB of execution-context device memory.

Demo

No demo video or Hugging Face Space exists for this product yet.

Limits

Reproducibility

Today an outsider cannot reproduce any of this hands-on: there is no public dataset, no demo Space, and the source is closed. The numbers above, including the scene-graph benchmark evaluation and the Jetson latency table, are reported from the project's own results files, not independently verifiable from outside yet.