JEV sees

JEV Sees adds a visual front end to JEV so cameras can provide structured decisions from images, video, and RGB-D streams.

JEV Sees connects images, video streams, and RGB-D cameras to the JEV judgment engine. It adds a visual front end that identifies objects, estimates risk, scores situations, and judges multiple visible objects at once. The system uses Florence-2 for dense-region captions and CLIP for color probability distributions, maintaining stable object IDs and scene memory across video frames. It is built for developers working on live video, robotics, monitoring, and testing systems who need repeated, structured decisions rather than open-ended VLM descriptions. The software requires Python 3.10+ and relies on the official typesafe-sdk for questions and responses.

JEV sees at a glance

Best for: Software engineer, Researcher
Runs on: macOS, Linux, Python
Pricing: Free
Released: September 2026
Category: AI research tool
Website: github.com/charlesfeng0314/jev_sees

Key features

Connects images, video streams, and RGB-D cameras to JEV
Maintains stable object IDs and scene memory across video frames
Preserves pixel measurements, CLIP color probability, and Florence region captions
Calculates metric 3D positions and gaps when RGB-D is available
Outputs program-ready choices, probabilities, and scores
Supports labeled bounding boxes and live video panels with save options

Use it for

Estimate accident-risk probability for visible pedestrians
Determine which object a robot should pay attention to
Track whether a target object has changed state
Judge many visible objects at once in a single call

Alternatives to JEV sees

Vev — Runs locally on NVIDIA hardware with a fast, Jev-compatible API.
NanoJev — It runs parallel decisions and outputs probability distributions without output-token decoding.
Kev — Provides calibrated probabilities and matches the TypeSafe API.
eu/jev — Inference hosted in Germany on Hetzner servers with no storage of prompts.
SemIf OpenJev — Extracts direct token probabilities without generating text or running JSON repair loops.
OpenVLHarness — Lifts six multimodal model backbones by 9.7 to 12.6 points across 22 benchmarks.
VLMEvalKit — Evaluate large vision-language models across numerous benchmarks with a single command.
OpenDecision — It uses SigLIP2 to quickly match images against candidate captions and rules.
Video Prediction Policy 2 — Achieves state-of-the-art performance on the RoboDojo-Sim leaderboard.
LiveFrame — It provides verified runtime, control, and stability reports with digest-bound ledgers.
Simo — Delivers calibrated probabilities and actionable values in milliseconds.
LMMs-Eval — Provides open research and tooling for multimodal model evaluation and benchmarks.

All JEV sees alternatives