Perception & computer vision

MOTHER DeepVision

The eyes of MOTHER — real-time detection, tracking and segmentation across every feed.

MOTHER DeepVision is our sovereign computer-vision pipeline and the visual cortex behind MOTHER EXO. It detects, tracks, segments and relates objects, faces and expressions across camera, drone and satellite imagery in real time — YOLOv8, RF-DETR and Segment Anything, trained on real data with TorchGeo for geospatial feeds. DeepVision is what lets MOTHER see.

SigLIP
SO400M ViT · 896px
kNN 1.0
measured retrieval
YOLOv8
RF-DETR · SAM ensemble
Real-time
detect · track · segment

Capabilities

Object detection

High-throughput detection with YOLOv8 and RF-DETR across camera, drone and satellite feeds.

Multi-object tracking

Persistent tracks across frames and cameras, with re-identification and motion estimation.

Segmentation

Pixel-accurate masks with the Segment Anything Model for precise scene understanding.

Faces & expression

Face detection and expression/affect reading that feed EXO's emotion understanding.

Relations & scene graphs

Builds structured relationships between detected entities — who is near, holding or interacting with what.

Geospatial vision

TorchGeo-powered analysis of satellite and aerial imagery for mapping and change detection.

Architecture

  1. 1

    Detector ensemble

    YOLOv8 for speed, RF-DETR for accuracy and SAM for segmentation — combined into one perception pipeline.

  2. 2

    Frozen ViT backbone

    A MOTHER DeepVision Vision Transformer provides the shared visual features that MOTHER EXO reasons over — trained once, reused everywhere.

  3. 3

    Training pipeline

    Python 3.10+ / PyTorch training with real datasets and TorchGeo for geospatial feeds; heads evaluated on real frames and clips.

Specifications

BackboneSigLIP-SO400M ViT · 896px · frozen (feeds EXO)
MeasuredkNN 1.0 on retrieval eval
DetectorsYOLOv8 · RF-DETR · Segment Anything (SAM)
Tasksdetect · track · segment · face · expression · relations
GeospatialTorchGeo (satellite / aerial)
ArmsVision (live) · T2V (placeholder) · Audio/Dialog (EXO heads)
ServingPython 3.10+ · PyTorch · sovereign on-node

Models & arms — function · dataset

WeightKindFunctionDatasetTrained
SigLIP-SO400M ViTbackboneSelf-supervised visual foundation (896px) — the shared features EXO reasons overall on-node imagery · kNN 1.0
YOLOv8detectorFast object detectionCOCO / LVIS
RF-DETRdetectorHigh-accuracy detectionCOCO / remote-sensing
SAMsegmenterPixel-accurate masksSegment Anything
face · expression · relationsheadsConsent face-ID · affect · scene-graph — on the backboneconsent gallery · affect · scene-graph
T2VarmText-to-video (latent diffusion)tokenizer TBDplaceholder
Audio / DialogarmASR + text-to-voice — lives in the EXO speech/audio headsLibriSpeech + codecin EXO

Model evaluation — industry benchmarks

BenchmarkMetricResultStatus
ImageNet kNNtop-11.0◉ measured
COCO detectionmAP50○ to run
ADE20k segmentationmIoU○ to run
LVISmAP○ to run
MOT17 trackingMOTA○ to run

Benchmarks scheduled on this build — measured results are recorded here as each run completes.

What it's for

Surveillance & ISR

Real-time detection and tracking across large camera estates and drone feeds for situational awareness.

Geospatial intelligence

Change detection and object analysis on satellite and aerial imagery with TorchGeo.

Perception for EXO

The visual cortex that supplies MOTHER EXO with the features it needs to build a world model.

Safety & governance

  • Observation and decision-support only — DeepVision informs, humans decide.
  • Runs on sovereign infrastructure with no data egress.
  • Evaluated on real imagery; performance is measured, never asserted.

Frequently asked

What is MOTHER DeepVision?

MOTHER DeepVision is Media Stream AI's sovereign computer-vision pipeline — real-time detection, tracking, segmentation, face and expression analysis and scene relations using YOLOv8, RF-DETR and Segment Anything, and the visual cortex behind MOTHER EXO.

What models does DeepVision use?

An ensemble of YOLOv8 (speed), RF-DETR (accuracy) and the Segment Anything Model (segmentation), plus a frozen DeepVision Vision Transformer backbone that feeds MOTHER EXO.

Can DeepVision analyse satellite imagery?

Yes — DeepVision uses TorchGeo for geospatial analysis of satellite and aerial imagery, including mapping and change detection.

How does DeepVision relate to MOTHER EXO?

DeepVision is EXO's eyes: its frozen ViT backbone provides the visual features EXO reasons over to build its world model.

The MOTHER model family

Build on MOTHER DeepVision

Sovereign, on-node and observe-and-advise by design. Talk to us about access and deployment.