English | 中文

JEV is fast, structured, and built to make decisions.

Connect an image, a video stream, or an RGB-D camera to JEV, and suddenly the same judgment engine can work on the visual world: identify what matters, estimate risk, score a situation, or judge many visible objects at once.

One frame. Many objects. One JEV call.

In the demo above, every visible pedestrian gets an accident-risk probability from the same JEV call.

JEV judges. JEV Sees lets it see. JEV Control lets it act.

JEV is already good at fast, closed, structured judgment: choice, probability, score.

But without vision, a huge class of useful questions is simply unavailable:

• Is this pedestrian in danger?
• Which object should the robot pay attention to?
• Has the target changed state?
• Which visible item best matches the condition?
• Can I ask the same question about every person in this frame?

JEV Sees opens that door.

It adds a visual front end to JEV so a camera can become another source of decisions — not just something that produces a description.

VLMs are great when you want open-ended visual understanding: describe a scene, answer a broad question, explain what is happening.

JEV is interesting when the question is already known and you want the answer fast, structured, repeatable, and easy to run many times.

That difference becomes especially useful in live video, robotics, monitoring, testing, and other systems where the same kind of decision may need to be made again and again.

Anything that can be expressed as a closed judgment over the scene.

from jev_sees import Choice, Sees, TypeSafeClient

question = "What color is the bus?"
sees = Sees()
sees.observe("assets/bus.jpg")

with TypeSafeClient() as client:
    response = client.system_one(
        state=sees.state(question),
        questions={
            "bus_color": Choice(
                instructions=question,
                criteria={"yellow": None, "red": None, "blue": None, "uncertain": None},
            )
        },
    )

print(response.choices["bus_color"].choice)
# blue

JEV Sees creates state. The official typesafe-sdk still owns the Choice, the system_one call, and the response.

from jev_sees import Noul

question = "Is object_003 in immediate danger from a vehicle?"
state = sees.state(question)

with TypeSafeClient() as client:
    response = client.system_one(
        state=state,
        questions={"in_danger": Noul(instructions=question)},
    )

print(response.nouls["in_danger"].noul)
# probability of "yes"

questions = {
    obj["object_id"]: Noul(
        instructions=f"Is {obj['object_id']} in immediate danger from a car?"
    )
    for obj in tracks
    if obj["label"] == "person"
}

with TypeSafeClient() as client:
    response = client.system_one(state=sees.state("Pedestrian risk"), questions=questions)

Each pedestrian remains a named official Noul question, and its official answer is available from response.nouls[object_id].

git clone https://github.com/CharlesFeng0314/JEV_sees.git
cd JEV_sees
python -m pip install -e .

JEV Sees requires Python 3.10+.

Create a key at console.typesafe.ai/keys.

macOS / Linux:

export TYPESAFE_API_KEY="your-key"

PowerShell:

$env:TYPESAFE_API_KEY = "your-key"

Both the official TypeSafeClient and the high-level video entry point can read this key. observe() and state() themselves remain local.

from jev_sees import Choice, Sees, TypeSafeClient

question = "What color is the bus?"
sees = Sees()
sees.observe("assets/bus.jpg")
with TypeSafeClient() as client:
    response = client.system_one(
        state=sees.state(question),
        questions={
            "bus_color": Choice(
                instructions=question,
                criteria={"yellow": None, "red": None, "blue": None, "uncertain": None},
            )
        },
    )
print(response.choices["bus_color"].choice)

Expected result on the included image:

blue

Full example: examples/bus_color.py

observe() is the local visual layer behind JEV Sees:

tracks = sees.observe("assets/bus.jpg")

for obj in tracks:
    print(obj["object_id"], obj["label"], obj["bbox_xyxy"])

A typical tracked object contains:

object_id
label
confidence
bbox_xyxy
centroid_uv
attributes.color_evidence.cv
attributes.color_evidence.clip
attributes.color_evidence.caption

Color is evidence, not an SDK verdict. The CV branch preserves pixel measurements, CLIP preserves its complete color probability distribution, and Florence's original region caption remains alongside both. JEV can therefore judge agreement or disagreement instead of receiving one preselected color string.

Across video frames, JEV Sees keeps object IDs stable when possible and maintains scene memory. Samples reuse the existing pose_history and add frame_index plus media time. The two most recent time-aware poses of each visible object enter the JEV state, so JEV can reason from box movement and the actual interval. Visual questions stay attached to the same object over time instead of treating every frame as a completely new world.

The scene can also carry useful spatial context such as:

• what is visible now
• what was seen earlier
• whether a remembered object is stale
• bounding-box overlap and gap
• centroid distance
• whether nearby objects appear to be approaching
• metric 3D positions and gaps when RGB-D is available

These are implementation details, but they unlock the product behavior that matters: JEV can keep making structured judgments about a changing visual scene.

Choice, Noul, Score, and TypeSafeClient are re-exported by jev_sees for a single import line. They are the official typesafe-sdk classes, so direct calls still use client.system_one(...) and official typed collections such as response.choices, response.nouls, and response.scores.

JEV Sees does not infer a question type from natural language and does not turn lists or dictionaries into JEV questions. Applications construct the official question objects themselves. For video, questions= may be a callable that receives the current tracked objects and returns a mapping of official questions.

examples/traffic_relations.py is intentionally a minimal terminal example. Its small questions() function is application code: it selects current pedestrians and creates an official Noul for each object ID. JEV Sees does not contain a traffic-specific plan or inspect the prompt to invent questions or options.

Video results are frame-primary in result.frames: every evaluated sample contains its frame_index, video_time_s, and the probability for every selected object in that frame. result.rows remains a per-object peak summary for compatibility. When save= is used, video output includes labeled bounding boxes and a right-side live panel whose object rows stay pinned after first detection.

Run it with:

python examples/traffic_relations.py

This example calls JEV and therefore requires TYPESAFE_API_KEY.

Source video credits: assets/CREDITS.md

RGB-only mode starts from Florence-2 dense-region captions. Florence discovers boxes and generates their labels; callers do not pass an object vocabulary.

RGB-D mode takes a different path: depth clusters decide which physical objects exist, and overlapping Florence regions provide free-text names when available.

That matters when a 2D detector misses something that is still physically present.

With camera intrinsics, RGB-D observations can also include position_m, which lets the scene state carry metric 3D positions and object-to-object gaps.

The earlier YOLO benchmark does not describe the Florence-2 pipeline and has intentionally been removed from the current documentation. A new image and video benchmark is required before publishing latency claims for this backend.

Weights are intentionally not stored in this repository.

Florence-2 is loaded from Hugging Face on first use. The default is microsoft/Florence-2-base-ft; choose another compatible checkpoint with:

Sees(florence_model="microsoft/Florence-2-large-ft")

If JEV_SEES_ROBO_ROOT points to a directory containing:

weights/clip/ViT-B-32.pt

JEV Sees uses those files.

JEV Sees uses that local CLIP weight for color evidence. Otherwise CLIP downloads its weight on first use.

If this fails:

python -c "import jev_sees"

the most common cause is that pip installed the package into a different Python environment from the python command you are using.

Use:

python -m pip install -e .
python -c "import jev_sees; print(jev_sees.__version__)"

Using python -m pip keeps installation and execution on the same interpreter.

Run the test suite with:

python -m unittest discover -s tests -v

JEV Sees is part of a simple idea: JEV should not stop at text.

Give it eyes, and it can judge the visual world. Give it hands, and those judgments can become actions.

              JEV
       structured judgment
          /           \
         /             \
   JEV Sees        JEV Control
      eyes             hands
     vision          robot action

• JEV Sees — give JEV visual input from images, video, and RGB-D cameras
• JEV Control Your Roboarm — use JEV judgments to choose robot-arm actions

The long-term idea is straightforward: see → judge → act.

JEV Sees is currently v0.1.0 and intentionally experimental.

The public surface stays small: Sees(...) is the high-level image/video product entry point, while observe() and state() expose the visual layer for direct official JEV calls. Official JEV classes are re-exported unchanged. The perception stack, scene representation, examples, and evaluation are still evolving.

If you try it on another camera, another robot, a weird scene, or a use case the examples did not anticipate, open an issue or start a discussion. Those experiments are exactly what this repo is for.

MIT