A Vertex AI Vision alternative for live cameras
Google's Vertex AI Vision reached end of life on September 30, 2026. Pictograph runs a detector you train on live video from RTSP cameras, browsers, and phones, and counts what crosses a line or stays in a zone.
Pictograph vs Vertex AI Vision
What each does with a live camera, from Google's documentation and Pictograph's.
| Feature | Pictograph | Vertex AI Vision |
|---|---|---|
| Available after September 30, 2026 | No: end of life | |
| Live video in | RTSP, WebRTC (WHIP / WHEP), files | RTSP or files, sent by vaictl |
| Connects to the camera itself | No: vaictl runs on its network | |
| Reconnects to a dropped camera | No: restart vaictl yourself | |
| What it detects | Any class you train | Prebuilt models, or custom ones |
| Line crossings, zones, dwell time | Workflows, on any class | Occupancy model: people, vehicles |
| Frames analyzed | Every frame the GPU keeps up with | 6 a second (occupancy model) |
| Events to your systems | Signed webhooks | Pub/Sub, through Cloud Functions |
| People counting with no training | – | |
| Person and face blur | – | |
| Searchable video archive | Event frames saved to a dataset | Vision Warehouse |
Where Pictograph fits
Camera projects that counted people or vehicles across a line, in a zone, or over time, and need a home now that the service is gone. You train the detector on your own footage, so it can count any class, not only people and vehicles. A workflow tracks, counts, and sends events on a GPU in us-east or us-west, and connects to the camera itself, with no ingestion tool to keep running.
Where Vertex AI Vision led
People and vehicle counting with no training, person and face blur, and Vision Warehouse, which stored video and made it searchable alongside the results, with output to BigQuery, all inside Google Cloud. If blurring people or archiving all your video is the requirement, Pictograph does not do either.
Stream a camera to a counting workflow
Once the workflow has a model and its lines, one command or a few lines of Python start the stream.
# One JSON line per frame, with each line's and zone's counts; Ctrl-C stops the stream
pictograph workflows stream "Gate counter" rtsp://camera.example.com:554/live --token pk_deploy_...from pictograph import Client
client = Client()
wf = client.workflows.get(
workflow="Gate counter",
)
live = client.workflows.connect(
workflow=wf,
token="pk_deploy_...",
)
with live.open_stream(source="rtsp://camera.example.com:554/live") as stream:
for frame in stream.frames():
print(frame.index, frame.measures)The workflow's endpoint pulls the RTSP camera itself; a browser or an app can push its own video over WebRTC instead. Each frame comes back with its detections and every line's and zone's running counts. The workflows reference covers lines, zones, and live events, and the streams guide covers every video source and the live view.
Move an occupancy app in four steps
Vertex AI Vision's occupancy model came pretrained; here the model is yours, so the first three steps build it.
-
Collect frames
Upload a few hundred frames from your cameras, or a short recording that Pictograph splits into frames.
-
Draft and review labels
SAM3 drafts every "person" or "car" from a text prompt. You review the drafts before training.
-
Train a detector
Export the reviewed dataset and train an RF-DETR or YOLOX detector on a managed GPU.
-
Count and stream
Start a workflow from the line counter, occupancy, or dwell time template, draw its lines and zones on a frame from your camera, and stream the camera to it.
The video API extracts frames, auto-annotation covers SAM3's prompts, the training guide walks through a run, and workflows covers tracking, counting, and webhooks.
Vertex AI Vision alternative FAQ
Is Vertex AI Vision shut down?
Yes. Google deprecated Vertex AI Vision on June 15, 2026, and it reached end of life on September 30, 2026. Google points to the Cloud Vision API, Agent Platform serverless training, and Agent Platform models as the way to migrate, which leaves getting video in from cameras, tracking, and counting for you to build.
Can Pictograph count people and vehicles like the occupancy analytics model?
Yes, once you train a detector for them. Vertex AI Vision's occupancy model came pretrained; on Pictograph the model is yours. Upload frames from your cameras, let SAM3 draft "person" and "car" labels from text prompts, review them, and train an RF-DETR or YOLOX detector. A workflow then tracks each object and counts it: a line counts crossings in each direction, and a zone counts entries and exits, the time spent inside, and visitors who stay a minimum time. Templates for a line counter, occupancy, and dwell time start you off.
How does camera video get to Pictograph without vaictl?
A workflow's endpoint connects to an RTSP camera itself and connects again if the camera drops, so nothing has to run on your network for a camera the cloud can reach. A camera that is only reachable inside your network can be pushed out over WebRTC (WHIP) from a machine there running GStreamer or OBS, and a browser or phone can stream its own camera. Each frame's results, with every line's and zone's counts, come back as the frame is processed.
How do I get counts and events into my systems?
Three ways. Read every frame's results and counts from the stream with the SDK or a WebSocket. Add a webhook output, which posts each event you pick (a line crossed, a zone entered or left, a dwell reached, a new object appearing) to your endpoint, signed and retried, at most one delivery a second with every event since. Or add an annotations output, which saves the event's frame to a dataset so you can review it or train on it.
What did Vertex AI Vision do that Pictograph does not?
People and vehicle counting with no training, person and face blur, and Vision Warehouse, which stored video and made it searchable alongside the model's results, with output to BigQuery. Pictograph does not blur people and does not archive video: it keeps the frames your events save, and the results it sends you.
Other comparisons
Count what your cameras see
$5 of free compute every month. No card required.